Omega: exible, scalable schedulers for large compute clusters

Malte Schwarzkopf

Andy Konwinski

Michael Abd-El-Malek

John Wilkes

University of Cambridge Computer Laboratory

University of California, Berkeley

Google, Inc.

[email protected]

[email protected]

{mabdelmalek,johnwilkes}@google.com
Abstract
Increasing scale and the need for rapid response to changing
requirements are hard to meet with current monolithic clus-
ter scheduler architectures. This restricts the rate at which
new features can be deployed, decreases efciency and uti-
lization, and will eventually limit cluster growth. We present
a novel approach to address these needs using parallelism,
shared state, and lock-free optimistic concurrency control.
We compare this approach to existing cluster scheduler
designs, evaluate how much interference between schedulers
occurs and how much it matters in practice, present some
techniques to alleviate it, and nally discuss a use case
highlighting the advantages of our approach all driven by
real-life Google production workloads.
Categories and Subject Descriptors D.4.7 [Operating
Systems]: Organization and DesignDistributed systems;
K.6.4 [Management of computing and information systems]:
System ManagementCentralization/decentralization
Keywords Cluster scheduling, optimistic concurrency con-
trol
1. Introduction
Large-scale compute clusters are expensive, so it is impor-
tant to use them well. Utilization and efciency can be in-
creased by running a mix of workloads on the same ma-
chines: CPU- and memory-intensive jobs, small and large
ones, and a mix of batch and low-latency jobs ones that
serve end user requests or provide infrastructure services
such as storage, naming or locking. This consolidation re-
duces the amount of hardware required for a workload, but
it makes the scheduling problem (assigning jobs to ma-
chines) more complicated: a wider range of requirements

Work done while interning at Google, Inc.

Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. To copy otherwise, to republish, to post on servers or to redistribute
to lists, requires prior specic permission and/or a fee.
EuroSys13 April 15-17, 2013, Prague, Czech Republic
Copyright c 2013 ACM 978-1-4503-1994-2/13/04. . . $15.00
subset
Monolithic Two-level Shared state
cluster
machines
cluster state
information
scheduling
logic
pessimistic
concurrency
(oers)
optimistic
concurrency
(transactions)
no
concurrency
full state
Figure 1: Schematic overview of the scheduling architec-
tures explored in this paper.
and policies have to be taken into account. Meanwhile, clus-
ters and their workloads keep growing, and since the sched-
ulers workload is roughly proportional to the cluster size,
the scheduler is at risk of becoming a scalability bottleneck.
Googles production job scheduler has experienced all
of this. Over the years, it has evolved into a complicated,
sophisticated system that is hard to change. As part of a
rewrite of this scheduler, we searched for a better approach.
We identied the two prevalent scheduler architectures
shown in Figure 1. Monolithic schedulers use a single,
centralized scheduling algorithm for all jobs (our existing
scheduler is one of these). Two-level schedulers have a sin-
gle active resource manager that offers compute resources to
multiple parallel, independent scheduler frameworks, as
in Mesos [13] and Hadoop-on-Demand [4].
Neither of these models satised our needs. Monolithic
schedulers do not make it easy to add new policies and spe-
cialized implementations, and may not scale up to the clus-
ter sizes we are planning for. Two-level scheduling archi-
tectures do appear to provide exibility and parallelism, but
in practice their conservative resource-visibility and locking
algorithms limit both, and make it hard to place difcult-
to-schedule picky jobs or to make decisions that require
access to the state of the entire cluster.
Our solution is a new parallel scheduler architecture built
around shared state, using lock-free optimistic concurrency
control, to achieve both implementation extensibility and
performance scalability. This architecture is being used in
351
Omega, Googles next-generation cluster management sys-
tem.
1.1 Contributions
The contributions of this paper are as follows. We:
1. present a lightweight taxonomy of the option space for
cluster scheduler development (3);
2. introduce a new scheduler architecture using shared state
and lock-free optimistic concurrency control (3.4);
3. compare the performance of monolithic, two-level and
shared-state scheduling using simulations and synthetic
workloads (4);
4. explore the behavior of the shared-state approach in more
detail using code based on a production scheduler and
driven by real-world workload traces (5); and
5. demonstrate the exibility of the shared-state approach
by means of a use case: we add a scheduler that uses
knowledge of the global cluster utilization to adjust the
resources given to running MapReduce jobs (6).
We nd that the Omega shared-state architecture can de-
liver performance competitive with or superior to other ar-
chitectures, and that interference in real-world settings is
low. The ability to access the entire cluster state in a sched-
uler brings other benets, too, and we demonstrate this by
showing how MapReduce jobs can be accelerated by using
spare resources.
2. Requirements
Cluster schedulers must meet a number of goals simulta-
neously: high resource utilization, user-supplied placement
constraints, rapid decision making, and various degrees of
fairness and business importance all while being robust
and always available. These requirements evolve over time,
and, in our experience, it becomes increasingly difcult to
add new policies to a single monolithic scheduler. This is
not just due to accumulation of code as functionality grows
over time, but also because some of our users have come
to rely on a detailed understanding of the internal behavior
of the system to get their work done, which makes both its
functionality and structure difcult to change.
2.1 Workload heterogeneity
One important driver of complexity is the hardware and
workload heterogeneity that is commonplace in large com-
pute clusters [24].
To demonstrate this, we examine the workload mix on
three Google production compute clusters that we believe to
be representative. Cluster A is a medium-sized, fairly busy
one, while cluster B is one of the larger clusters currently
in use at Google, and cluster C is the one for which a
scheduler workload trace was recently published [24, 27].
The workloads are from May 2011. All the clusters run a
J J J T T T C C C R R R
0.0
0.2
0.4
0.6
0.8
1.0
A B C Batch Service
Figure 2: Batch and service workloads for the clusters A,
B, and C: normalized numbers of jobs (J) and tasks (T), and
aggregate requests for CPU-core-seconds (C) and RAMGB-
seconds (R). The striped portion is the service jobs; the rest
is batch jobs.
1s 1min 1h 1d 29d
Job runtime [log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
A
B
C
1ms 1s 1min 1h
Interarrival time [log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
A
B
C
Figure 3: Cumulative distribution functions (CDFs) of job
runtime and job inter-arrival times for clusters A, B, and C.
Where the lines do not meet 1.0, some of the jobs ran for
longer than the 30-day range. In this and subsequent graphs,
solid lines represent batch jobs, and dashed lines are for
service jobs.
wide variety of jobs; some are congured by hand; some by
automated systems such as MapReduce [8], Pregel [19] and
Percolator [23].
There are many ways of partitioning a clusters workload
between different schedulers. Here, we pick a simple two-
way split between long-running service jobs that provide
end-user operations (e.g., web services) and internal infras-
tructure services (e.g., BigTable [5]), and batch jobs which
perform a computation and then nish. Although many other
splits are possible, for simplicity we put all lowpriority jobs
1
and those marked as best effort or batch into the batch
category, and the rest into the service category.
A job is made up of one or more tasks sometimes
thousands of tasks. Most (>80%) jobs are batch jobs, but the
majority of resources (5580%) are allocated to service jobs
(Figure 2); the latter typically run for much longer (Figure
3), and have fewer tasks than batch jobs (Figure 4). These
results are broadly similar to other analyses of cluster traces
from Yahoo [17], Facebook [7] and Google [20, 24, 25, 29].
Why does this matter? Many batch jobs are short, and
fast turnaround is important, so a lightweight, low-quality
1
In the public trace for cluster C, these are priority bands 08 [27].
352
1 10 100 1k 10k
Number of tasks [log
10
]
0.5
0.6
0.7
0.8
0.9
1.0
A
B
C
100 1k 10k
Number of tasks [log
10
]
0.95
0.96
0.97
0.98
0.99
1.00
A
B
C
Figure 4: CDF of the number of tasks in a job for clusters A,
B, and C. The right hand graph is an expansion of the tail of
the left-hand one, looking at 95
th
percentile, 100 tasks.
approach to placement works just ne. But long-running,
high-priority service jobs (2040% of them run for over a
month) must meet stringent availability and performance tar-
gets, meaning that careful placement of their tasks is needed
to maximize resistance to failures and provide good per-
formance. Indeed, the Omega service scheduler will try to
place tasks to resist both independent and coordinated fail-
ures, which is an NP-hard chance-constrained optimization
problem with tens of failure domains that nest and overlap.
Our previous implementation could take tens of seconds to
do this. While it is very reasonable to spend a few seconds
making a decision whose effects last for several weeks, it
can be problematic if an interactive batch job has to wait for
such a calculation. This problem is typically referred to as
head of line blocking, and can be avoided by introducing
parallelism.
In summary, what we require is a scheduler architec-
ture that can accommodate both types of jobs, exibly sup-
port job-specic policies, and also scale to an ever-growing
amount of scheduling work. The next section examines some
of these requirements in greater detail, as well as some ap-
proaches to meeting them.
3. Taxonomy
We begin with a short survey of the design issues cluster
schedulers must address, followed by an examination of
some different scheduler architectures that might meet them.
Partitioning the scheduling work. Work can be spread
across schedulers by (1) load-balancing that is oblivious
to workload type; (2) dedicating specialized schedulers to
different parts of the workload; or (3) a combination of the
two. Some systems use multiple job queues to hold the job
requests (e.g., for different priorities), but that does not affect
the scheduling parallelism: we are more interested in how
many schedulers are assigned to process the queues.
Choice of resources. Schedulers can be allowed to se-
lect from all of the cluster resources, or limited to a subset
to streamline decision making. The former increases the op-
portunity to make better decisions, and is important when
picky jobs need to be placed into a nearly-full cluster, or
when decisions rely on overall state, such as the total amount
of unused resources. Schedulers can have greater exibility
in placing tasks if they can preempt existing assignments, as
opposed to merely considering idle resources, but this comes
at the cost of wasting some work in the preempted tasks.
Interference. If schedulers compete for resources, mul-
tiple schedulers may attempt to claim the same resource si-
multaneously. Apessimistic approach avoids the issue by en-
suring that a particular resource is only made available to one
scheduler at a time; an optimistic one detects the (hopefully
rare) conicts, and undoes one or more of the conicting
claims. The optimistic approach increases parallelism, but
potentially increases the amount of wasted scheduling work
if conicts occur too frequently.
Allocation granularity. Since jobs typically contain
many tasks, schedulers can have different policies for how
to schedule them: at one extreme is atomic all-or-nothing
gang scheduling of the tasks in a job, at the other is incre-
mental placement of tasks as resources are found for them.
An all-or-nothing policy can be approximated by incremen-
tally acquiring resources and hoarding them until the job
can be started, at the cost of wasting those resources in the
meantime.
All have downsides: gang scheduling may be needed by
some jobs (e.g., MPI programs), but can unnecessarily delay
the start of others that can make progress with only a fraction
of their requested resources (e.g., MapReduce jobs). Incre-
mental resource acquisition can lead to deadlock if no back-
off mechanism is provided, while hoarding reduces cluster
utilization and can also cause deadlock.
Cluster-wide behaviors. Some behaviors span multi-
ple schedulers. Examples include achieving various types
of fairness, and a common agreement on the relative im-
portance of work, especially if one scheduler can preempt
others tasks. Strict enforcement of these behaviors can be
achieved with centralized control, but it is also possible to
rely on emergent behaviors to approximate the desired be-
havior. Techniques such as limiting the range of priorities
that a scheduler can wield can provide partial enforcement of
desired behaviors, and compliance to cluster-wide policies
can be audited post facto to eliminate the need for checks in
a schedulers critical code path.
This space is obviously larger than can be explored in a
single paper; we focus on the combinations that are summa-
rized in Table 1, and described in greater detail in the next
few sections.
3.1 Monolithic schedulers
Our baseline for comparisons is a monolithic scheduler that
has but a single instance, no parallelism, and must imple-
ment all the policy choices in a single code base. This
approach is common in the high-performance computing
(HPC) world, where a monolithic scheduler usually runs a
353
Approach Resource choice Interference Alloc. granularity Cluster-wide policies
Monolithic all available none (serialized) global policy strict priority (preemption)
Statically partitioned xed subset none (partitioned) per-partition policy scheduler-dependent
Two-level (Mesos) dynamic subset pessimistic hoarding strict fairness
Shared-state (Omega) all available optimistic per-scheduler policy free-for-all, priority preemption
Table 1: Comparison of parallelized cluster scheduling approaches.
single instance of the scheduling code, and applies the same
algorithm for all incoming jobs. HPC schedulers such as
Maui [16] and its successor Moab, as well as Platform LSF
[14], support different policies by means of a complicated
calculation involving multiple weighting factors to calculate
an overall priority, after which the scheduler can roughly
fulll site objectives by starting the jobs in priority order
[1].
Another way to support different scheduling policies is to
provide multiple code paths in the scheduler, running sep-
arate scheduling logic for different job types. But this is
harder than it might appear. Googles current cluster sched-
uler is effectively monolithic, although it has acquired many-
optimizations over the years to provide internal parallelism
and multi-threading to address head-of-line blocking and
scalability. This complicates an already difcult job: the
scheduler has to minimize the time a job spends waiting
before it starts running, while respecting priorities, per-job
constraints [20, 25], and a number of other policy goals such
as failure-tolerance and scaling to workloads that ll many
thousands of machines. Although it has been hugely success-
ful, our scheduler has experienced several years of evolution
and organic software growth, and we have found that it is
surprisingly difcult to support a wide range of policies in
a sustainable manner using a single-algorithm implementa-
tion. In the end, this kind of software engineering consider-
ation, rather than performance scalability, was our primary
motivation to move to an architecture that supported concur-
rent, independent scheduling components.
3.2 Statically partitioned schedulers
Most cloud computing schedulers (e.g., Hadoop [28], and
Dryads Quincy [15]) assume they have complete control
over a set of resources, as they are typically deployed onto
dedicated, statically-partitioned clusters of machines; or by
partitioning a single cluster into different parts that support
different behaviors [6]. This leads to fragmentation and sub-
optimal utilization, which is not viable for us, and so we did
not explore this option any further.
3.3 Two-level scheduling
An obvious x to the issues of static partitioning is to ad-
just the allocation of resources to each scheduler dynami-
cally, using a central coordinator to decide how many re-
sources each sub-cluster can have. This two-level scheduling
approach is used by a number of systems, including Mesos
[13] and Hadoop-on-Demand (HOD) [4].
In Mesos, a centralized resource allocator dynamically
partitions a cluster, allocating resources to different sched-
uler frameworks.
2
Resources are distributed to the frame-
works in the form of offers, which contain only available
resources ones that are currently unused. The allocator
avoids conicts by only offering a given resource to one
framework at a time, and attempts to achieve dominant re-
source fairness (DRF) [11] by choosing the order and the
sizes of its offers.
3
Because only one framework is examin-
ing a resource at a time, it effectively holds a lock on that
resource for the duration of a scheduling decision. In other
words, concurrency control is pessimistic.
Mesos works best when tasks are short-lived and relin-
quish resources frequently, and when job sizes are small
compared to the size of the cluster. As we explained in 2.1,
our cluster workloads do not have these properties, espe-
cially in the case of service jobs, and 4.2 will show that this
makes an offer-based two-level scheduling approach unsuit-
able for our needs.
While a Mesos framework can use lters to describe
the kinds of resources that it would like to be offered, it does
not have access to a view of the overall cluster state just the
resources it has been offered. As a result, it cannot support
preemption or policies requiring access to the whole cluster
state: a framework simply does not have any knowledge
of resources that have been allocated to other schedulers.
Mesos uses resource hoarding to achieve gang scheduling,
and can potentially deadlock as a result.
It might appear that YARN [21] is a two-level scheduler,
too. In YARN, resource requests from per-job application
masters are sent to a single global scheduler in the resource
master, which allocates resources on various machines, sub-
ject to application-specied constraints. But the application
masters provide job-management services, not scheduling,
so YARN is effectively a monolithic scheduler architecture.
At the time of writing, YARN only supports one resource
type (xed-sized memory chunks). Our experience suggests
that it will eventually need a rich API to the resource master
2
We describe the most recently released version of Mesos at the time we
did this work: 0.9.0-incubating from May 8, 2012.
3
The Mesos simple allocator offers all available resources to a framework
every time it makes an offer, and does not limit the amount of resources
that a framework can accept. This negatively impacts Mesos as framework
decision times grow; see 4.2.
354
in order to cater for diverse application requirements, includ-
ing multiple resource dimensions, constraints, and place-
ment choices for failure-tolerance. Although YARN appli-
cation masters can request resources on particular machines,
it is unclear how they acquire and maintain the state needed
to make such placement decisions.
3.4 Shared-state scheduling
The alternative used by Omega is the shared state approach:
we grant each scheduler full access to the entire cluster, al-
low them to compete in a free-for-all manner, and use op-
timistic concurrency control to mediate clashes when they
update the cluster state. This immediately eliminates two
of the issues of the two-level scheduler approach limited
parallelism due to pessimistic concurrency control, and re-
stricted visibility of resources in a scheduler framework at
the potential cost of redoing work when the optimistic con-
currency assumptions are incorrect. Exploring this tradeoff
is the primary purpose of this paper.
There is no central resource allocator in Omega; all of the
resource-allocation decisions take place in the schedulers.
We maintain a resilient master copy of the resource alloca-
tions in the cluster, which we call cell state.
4
Each scheduler
is given a private, local, frequently-updated copy of cell state
that it uses for making scheduling decisions. The scheduler
can see the entire state of the cell and has complete free-
dom to lay claim to any available cluster resources provided
it has the appropriate permissions and priority even ones
that another scheduler has already acquired. Once a sched-
uler makes a placement decision, it updates the shared copy
of cell state in an atomic commit. At most one such com-
mit will succeed in the case of conict: effectively, the time
from state synchronization to the commit attempt is a trans-
action. Whether or not the transaction succeeds, the sched-
uler resyncs its local copy of cell state afterwards and, if
necessary, re-runs its scheduling algorithm and tries again.
Omega schedulers operate completely in parallel and do
not have to wait for jobs in other schedulers, and there is no
inter-scheduler head of line blocking. To prevent conicts
from causing starvation, Omega schedulers typically choose
to use incremental transactions, which accept all but the
conicting changes (i.e., the transaction provides atomicity
but not independence). A scheduler can instead use an all-
or-nothing transaction to achieve gang scheduling: either all
tasks of a job are scheduled together, or none are, and the
scheduler must try to schedule the entire job again. This
helps to avoid resource hoarding, since a gang-scheduled job
can preempt lower-priority tasks once sufcient resources
are available and its transaction commits, and allow other
schedulers jobs to use the resources in the meantime.
Different Omega schedulers can implement different
policies, but all must agree on what resource allocations are
4
A cell is the management unit for part of a physical cluster; a cluster may
support more than one cell. Cells do not overlap.
Lightweight (4) High-delity (5)
Machines homogeneous actual data
Resource req. size sampled actual data
Initial cell state sampled actual data
tasks per job sampled actual data

jobs
sampled actual data
Task duration sampled actual data
Sched. constraints ignored obeyed
Sched. algorithm randomized rst t Google algorithm
Runtime fast (24h 5 min.) slow (24h 2h)
Table 2: Comparison of the two simulators; actual data
refers to use of information found in a detailed workload-
execution trace taken from a production cluster.
permitted (e.g., a common notion of whether a machine is
full), and a common scale for expressing the relative impor-
tance of jobs, called precedence. These rules are deliberately
kept to a minimum. The two-level schemes centralized re-
source allocator component is thus simplied to a persistent
data store with validation code that enforces these common
rules. Since there is no central policy-enforcement engine
for high-level cluster-wide goals, we rely on these showing
up as emergent behaviors that result from the decisions of
individual schedulers. In this, it helps that fairness is not a
primary concern in our environment: we are driven more by
the need to meet business requirements. In support of these,
individual schedulers have conguration settings to limit the
total amount of resources they may claim, and to limit the
number of jobs they admit. Finally, we also rely on post-
facto enforcement, since we are monitoring the systems
behavior anyway.
The performance viability of the shared-state approach
is ultimately determined by the frequency at which transac-
tions fail and the costs of such failures. The rest of this paper
explores these issues for typical cluster workloads at Google.
4. Design comparisons
To understand the tradeoffs between the different approaches
described before (monolithic, two-level and shared-state
schedulers), we built two simulators:
1. A lightweight simulator driven by synthetic work-
loads using parameters drawn from empirical workload dis-
tributions. We use this to compare the behaviour of all three
architectures under the same conditions and with identical
workloads. By making some simplications, this lightweight
simulator allows us to sweep across a broad range of operat-
ing points within a reasonable runtime. The lightweight sim-
ulator also does not contain any proprietary Google code and
is available as open source software.
5
2. A high-delity simulator that replays historic work-
load traces from Google production clusters, and reuses
much of the Google production schedulers code. This gives
5
https://code.google.com/p/cluster-scheduler-simulator/.
355
10ms 0.1s 1s 10s 100s
t
job
[sec; log
10
]
0.1s
1s
1m
1h
1d
M
e
a
n
j
o
b
w
a
i
t
t
i
m
e
[
l
o
g
1
0
]
A
B
C
(a) Single-path.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.1s
1s
1m
1h
1d
M
e
a
n
j
o
b
w
a
i
t
t
i
m
e
[
l
o
g
1
0
]
A
B
C
(b) Multi-path.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.1s
1s
1m
1h
1d
M
e
a
n
j
o
b
w
a
i
t
t
i
m
e
[
l
o
g
1
0
]
A
B
C
(c) Shared state.
Figure 5: Schedulers job wait time, as a function of t
job
in the monolithic single-path case, t
job
(service) in the monolithic
multi-path and shared-state cases. The SLO (horizontal bar) is 30s.
10ms 0.1s 1s 10s 100s
t
job
[sec; log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
A
B
C
(a) Single-path.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
A
B
C
(b) Multi-path.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
A
B
C
(c) Shared state.
Figure 6: Schedulers busyness, as a function of t
job
in the monolithic single-path case, t
job
(service) in the monolithic multi-
path and shared-state cases. The value is the median daily busyness over the 7-day experiment, and error bars are one median
absolute deviation (MAD), i.e. the median deviation from the median value, a robust estimator of typical value dispersion.
us behavior closer to the real system, at the price of only
supporting the Omega architecture and running a lot more
slowly than the lightweight simulator: a single run can take
days.
The rest of this section describes the simulators and our
experimental setup.
Simplications in the lightweight simulator. In the
lightweight simulator, we trade speed and exibility for ac-
curacy by making some simplifying assumptions, summa-
rized in Table 2.
The simulator is driven by a workload derived from from
real workloads that ran on the same clusters and time periods
discussed in 2.1. While the high-delity simulator is driven
by the actual workload traces, for the lightweight simulator
we analyze the workloads to obtain distributions of parame-
ter values such as the number of tasks per job, the task du-
ration, the per-task resources and job inter-arrival times, and
then synthesize jobs and tasks that conform to these distri-
butions.
At the start of a simulation, the lightweight simulator ini-
tializes cluster state using task-size data extracted from the
relevant trace, but only instantiates sufciently many tasks
to utilize about 60% of cluster resources, which is compara-
ble to the utilization level described in [24]. In production,
Google speculatively over-commits resources, but the mech-
anisms and policies for this are too complicated to be repli-
cated in the lightweight simulator.
The simulator can support multiple scheduler types, but
initially we consider just two: batch and service. The two
types of job have different parameter distributions, summa-
rized in 2.1.
To improve simulation runtime in pathological situations,
we limit any single job to 1,000 scheduling attempts, and the
simulator abandons the job at this point if some tasks are still
unscheduled. In practice, this only matters for the two-level
scheduler (see 4.2), and is rarely triggered by the others.
Parameters. We model the scheduler decision time
as a linear function of the form t
decision
= t
job
+ t
task

tasks per job, where t

job
is a per-job overhead and t
task
rep-
resents the incremental cost to place each task. This turns
out to be a reasonable approximation of Googles current
cluster scheduling logic because most jobs in our real-life
356
workloads have tasks with identical requirements [24]. Our
values for t
job
and t
task
are based on somewhat conservative
6
estimates from measurements of our current production sys-
tems behavior: t
job
= 0.1s and t
task
= 5ms.
Many of our experiments explore the effects of vary-
ing t
decision
(service) for the service scheduler because we
are interested in exploring how Omega is affected by the
longer decision times needed for sophisticated placement al-
gorithms. We also vary the job arrival rate,
jobs
, to model
changes to the cluster workload level.
Metrics. Typically, users evaluate the perceived quality
of cluster scheduling by considering the time until their jobs
start running, as well as their runtime to completion. We re-
fer to the former metric as job wait time, which we dene as
the difference between the job submission time and the be-
ginning of the jobs rst scheduling attempt. Our schedulers
process one request at a time, so a busy scheduler will cause
enqueued jobs to be delayed. Job wait time thus measures the
depth of scheduler queues, and will increase as the scheduler
is busy for longer either because it receives more jobs, or
because they take longer to schedule. A common production
service level objective (SLO) for job wait time is 30s.
Job wait time depends on the scheduler busyness: the
fraction of time in which the scheduler is busy making
scheduling decisions. It increases with the per-job decision
time, and, in the shared-state approach, if scheduling work
must be redone because of conicts. To assess how much
of the latter is occurring, we measure the conict fraction,
which denotes the average number of conicts per success-
ful transaction. A value of 0 means no conicts took place;
a value of 3 indicates that the average job experiences three
conicts, and thus requires four scheduling attempts.
Our values for scheduler busyness and conict fraction
are medians of the daily values, and wait time values are
overall averages. Where present, error bars indicate how
much variation exists across days in the experiment: they
show the median absolute deviation (MAD) from the me-
dian value of the per-day averages. All experiments simulate
seven days of cluster operation, except for the Mesos ones,
which simulate only one day, as they take much longer to run
because of the failed scheduling attempts that result from in-
sufcient available resources (see 4.2).
4.1 Monolithic schedulers
Our baseline for comparison is a serial monolithic scheduler
with the same decision time for batch and service jobs, to
reect the need to run much of the same code for every
job type (a single-path implementation). We also consider
a monolithic scheduler with a fast code path for batch jobs;
we refer to this as a multi-path monolithic scheduler, since it
still schedules only one job at a time. The current monolithic
Google cluster scheduler is somewhere in between these
6
In the sense that they are approximations least favorable to the Omega
architecture.
pure designs: it does run some job-specic logic, but mostly
applies identical scheduling logic for all jobs.
In the baseline case, we vary the scheduler decision time
on the x-axis by changing t
job
. In the multi-path case, we
split the workload into batch and service workloads and use
the defaults for the batch scheduler decision time while we
vary t
job
(service).
The results are not surprising: in the single-path baseline
case, the scheduler busyness is low as long as scheduling is
quick, but scales linearly with increased t
job
(Figure 6a). As
a consequence, job wait time increases at a similar rate until
the scheduler is saturated, at which point it cannot keep up
with the incoming workload any more. The wait time curves
for service jobs closely track the ones for batch jobs, since
all jobs take the same time to schedule (Figure 5a).
With a fast path for batch jobs in the multi-path case,
both average job wait time and scheduler busyness decrease
signicantly even at long decision times for service jobs,
since the majority of jobs are batch ones. But batch jobs
can still get stuck in a queue behind the slow-to-schedule
service jobs, and head-of-line blocking occurs: scalability is
still limited by the processing capacity of a single scheduler
(Figures 5b and 6b). To avoid this, we need some form of
parallel processing.
4.2 Two-level scheduling: Mesos
Our two-level scheduler experiments are modeled on the
offer-based Mesos design. We simulate a single resource
manager and two scheduler frameworks, one handling batch
jobs and one handling service jobs. To keep things simple,
we assume that a scheduler only looks at the set of resources
available to it when it begins a scheduling attempt for a job
(i.e., any offers that arrive during the attempt are ignored).
Resources not used at the end of scheduling a job are re-
turned to the allocator; they may be re-offered again if the
framework is the one furthest below its fair share. The DRF
algorithm used by Mesoss centralized resource allocator is
quite fast, so we assume it takes 1 ms to make a resource
offer.
Since we now have two schedulers, we keep the deci-
sion time for the batch scheduler constant, and vary the deci-
sion time for the service scheduler by adjusting t
job
(service).
However, the batch scheduler busyness (Figure 7b) turns out
to be much higher than in the monolithic multi-path case.
This is a consequence of an interaction between the Mesos
offer model and the service schedulers long scheduling de-
cision times. Mesos achieves fairness by alternately offer-
ing all available cluster resources to different schedulers,
predicated on assumptions that resources become available
frequently and scheduler decisions are quick. As a result,
a long scheduler decision time means that nearly all clus-
ter resources are locked down for a long time, inaccessible
to other schedulers. The only resources available for other
schedulers in this situation are the few becoming available
while the slow scheduler is busy. These are often insufcient
357
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.1s
1s
1m
1h
1d
M
e
a
n
j
o
b
w
a
i
t
t
i
m
e
[
l
o
g
1
0
]
A
B
C
(a) Job wait time.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
A
B
C
(b) Scheduler busyness.
10ms 0.1s 1s 10s 100s
t
job
(service) [sec; log
10
]
0
40
80
120
160
200
J
o
b
s
a
b
a
n
d
o
n
e
d
A
B
C
(c) Unscheduled jobs.
Figure 7: Two-level scheduling (Mesos): performance as a function of t
job
(service).
to schedule an above-average size batch job, meaning that
the batch scheduler cannot make progress while the service
scheduler holds an offer. It nonetheless keeps trying, and
as a consequence, we nd that a number of jobs are aban-
doned because they did not nish scheduling their tasks by
the 1,000-attempt retry limit in the Mesos case (Figure 7c).
This pathology occurs because of Mesoss assumption
of quick scheduling decisions, small jobs and high re-
source churn, which do not hold for our service jobs. Mesos
could be extended to make only fair-share offers, although
this would complicate the resource allocator logic, and the
quality of the placement decisions for big or picky jobs
would likely decrease, since each scheduler could only see
a smaller fraction of the available resources. We have raised
this point with the Mesos team; they agree about the limita-
tion and are considering to address it in future work.
4.3 Shared-state scheduling: Omega
Finally, we use the lightweight simulator to explore the
Omega shared-state approach. We again simulate two sched-
ulers: one handling the batch workload, one handling the
service workload. Both schedulers refresh their local copy
of cell state by synchronizing it with the shared one when
they start looking at a job, and work on their local copy for
the duration of the decision time. Assuming at least one task
got scheduled, a transaction to update the shared cell state is
issued once nished. If there are no conicts, then the entire
transaction is accepted; otherwise only those changes that
do not result in an overcommitted machine are accepted.
Figure 5c shows that the average job wait times for the
Omega approach are comparable to those for multi-path
monolithic (Figure 5b). This suggests that conicts and in-
terference are relatively rare, and this is conrmed by the
graph of scheduler busyness (Figure 6c). Unlike Mesos (Fig-
ure 7c), the Omega-style scheduler manages to schedule all
jobs in the workload. Unlike the monolithic multi-path im-
plementation, it does not suffer from head-of-line blocking:
the lines for batch and service jobs are independent.
We also investigate at how the Omega approach scales
as the workload changes. For this purpose, we increase the
job arrival rate of the batch scheduler,
jobs
(batch). Figure
8 shows that both job wait time and scheduler busyness
increase. In the batch case, this is due to the higher job arrival
rate, while in the service case, it is due to additional conicts.
As indicated by the dashed vertical lines, cluster A scales to
about 2.5 the original workload before failing to keep up,
while clusters B and C scale to 6 and 9.5, respectively.
Since the batch scheduler is the main scalability bottle-
neck, we repeat the same scaling experiment with multiple
batch schedulers in order to test the ability of the Omega
model to scale to larger loads. The batch scheduling work
is load-balanced across the schedulers using a simple hash-
ing function. As expected, the conict fraction increases
with more schedulers as more opportunities for conict exist
(Figure 9a), but this is compensated at least up to 32 batch
schedulers by the better per-scheduler busyness with more
schedulers (Figure 9b). Similar results are seen with the job
wait times (not shown here). This is an encouraging result:
the Omega model can scale to a high batch workload while
still providing good behavior for service jobs.
4.4 Summary
The lightweight simulator is a useful tool for comparing the
different scheduler architectures. Figure 10 summarizes the
results graphically, considering the impact of scaling t
task
as
an additional dimension.
In short, the monolithic scheduler is not scalable. Al-
though adding the multi-path feature reduces the average
scheduling decision time, head-of-line blocking is still a
problem for batch jobs, and means that this model may not
be able to scale to the workloads we project for large clus-
ters. The two-level model of Mesos can support independent
scheduler implementations, but it is hampered by pessimistic
locking, does not handle long decision times well, and could
not schedule much of the heterogeneous load we offered it.
The shared-state Omega approach seems to offer compet-
itive, scalable performance with little interference at realistic
358
1x 2x 4x 6x 8x 10x
Relative
jobs
(batch)
0.1s
1s
1m
1h
1d
M
e
a
n
j
o
b
w
a
i
t
t
i
m
e
[
l
o
g
1
0
]
A
B
C
(a) Job wait time.
1x 2x 4x 6x 8x 10x
Relative
jobs
(batch)
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
A
B
C
(b) Scheduler busyness.
Figure 8: Shared-state scheduling (Omega): varying the ar-
rival rate for the batch workload,
jobs
(batch), for cluster B.
Dashed vertical lines indicate points of scheduler saturation;
i.e., only partial scheduling of the workload to their right.
operating points, supports independent scheduler implemen-
tations, and exposes the entire allocation state to the sched-
ulers. We show how this is helpful in 6. Our results indicate
that Omega can scale to many schedulers, as well as to chal-
lenging workloads.
5. Trace-driven simulation
Having compared the different scheduler architectures using
the lightweight simulator, we use the high-delity simulator
to explore some of the properties of the Omega shared-state
approach in greater detail and without the simplifying as-
sumptions made by the lightweight simulator. The core of
the high-delity simulator is the code used in Googles pro-
duction scheduling system. It respects task placement con-
straints, uses the same algorithms as the production ver-
sion, and can be given initial cell descriptions and detailed
workload traces obtained from live production cells. It lets
us evaluate the shared-state design with high condence on
1x 2x 4x 6x 8x 10x
Relative
jobs
(batch)
0.0
0.2
0.4
0.6
0.8
1.0
M
e
a
n
c
o
n

i
c
t
f
r
a
c
t
i
o
n
1
2
4
8
16
32
(a) Mean conict fraction.
1x 2x 4x 6x 8x 10x
Relative
jobs
(batch)
0.0
0.2
0.4
0.6
0.8
1.0
M
e
a
n
s
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
1
2
4
8
16
32
(b) Mean sched. busyness.
Figure 9: Shared-state scheduling (Omega): varying the ar-
rival rate for the batch workload (
jobs
(batch)) for cluster B;
1.0 is the default rate. Each line represents a different num-
ber of batch schedulers.
real-world workloads. We use it to answer the following
questions:
1. How much scheduling interference is present in real-
world workloads and what scheduler decision times can
we afford in production (5.1)?
2. What are the effects of different conict detection and
resolution techniques on real workloads (5.2)?
3. Can we take advantage of having access to the entire state
of the cell in a scheduler? (6)
Large-scale production systems are enormously compli-
cated, and thus even the high-delity simulator employs a
few simplications. It does not model machine failures (as
these only generate a small load on the scheduler); it does not
model the disparity between resource requests and the actual
usage of those resources in the traces (further discussed else-
where [24]); it xes the allocations at the initially-requested
sizes (a consequence of limitations in the trace data); and it
359
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0.001
100
10
1.0
0.1
(a) Monolithic scheduler,
single-path.
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0 001
100
10
1.0
0.1
(b) Monolithic scheduler,
multi-path.
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0.001
100
10
1.0
0.1
(c) Two-level scheduling
(Mesos).
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0 001
100
10
1.0
0.1
(d) Shared-state (Omega).
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0 001
100
10
1.0
0.1
(e) Shared-state, coarse,
gang scheduling.
Figure 10: Lightweight simulator: impact of varying t
job
(service) (right axis) and t
task
(service) (left axis) on scheduler busyness
(z-axis) in different scheduling schemes, on cluster B. Red shading of a 3D graph means that part of the workload remained
unscheduled.
0.4
1.0
0.8
0.6
0.2
0.0
1.0
0.1
0.01
0.001
100
10
1.0
0.1
t (se
rv
ice
)
t (service)
s
e
r
v
i
c
e
s
c
h
e
d
u
l
e
r

b
u
s
y
n
e
s
s
task
job
Figure 11: Shared-state scheduling (Omega): effect on
service scheduler busyness of varying t
job
(service) and
t
task
(service), using the high-delity simulator and a 29-day
trace from cluster C.
disables preemptions, because we found that they make lit-
tle difference to the results, but signicantly slow down the
simulations.
As expected, the outputs of the two simulators generally
agree. The main difference is that the lightweight simula-
tor runs experience less interference, which is likely a result
of the lightweight simulators lack of support for placement
constraints (which makes picky jobs seem easier to sched-
ule than they are), and its simpler notion of when a machine
is considered full (which means it sees fewer conicts with
ne-grained conict detection, cf. 5.2).
We can nonetheless conrm all the trends the lightweight
simulator demonstrates for the Omega shared-state model
using the high-delity simulator. We believe this conrms
that the lightweight simulator experiments provide plausible
comparisons between different scheduling architectures un-
der a common set of assumptions.
5.1 Scheduling performance
Figure 11 shows how service scheduler busyness varies as a
function of both t
job
(service) and t
task
(service) for a month-
long trace of cluster C (covering the same workload as the
public trace). Encouragingly, the scheduler busyness re-
mains low across almost the entire range for both, which
means that the Omega architecture scales well to long deci-
sion times for service jobs.
Scaling the workload. We also investigate the perfor-
mance of the shared-state architecture using a 7-day trace
from cluster B, which is one of the largest and busiest
Google clusters. Again, we vary t
job
(service). In Figure 12b,
once t
job
(service) reaches about 10s, the conict fraction in-
creases beyond 1.0, so that scheduling a service job requires
at least one retry, on average.
At around the same point, we fail to meet the 30s job
wait time SLO for the service scheduler (Figure 12a), even
though the scheduler itself is not yet saturated: the additional
wait time is purely due to the impact of conicts. To con-
rm this, we approximate the time that the scheduler would
have taken if it had experienced no conicts or retries (the
no conict case in Figure 12c), and nd that the service
scheduler busyness with conicts is about 40% higher than
in the no-conict case. This is a higher level of interference
compared to cluster C, most likely because of a much higher
batch load in cluster B.
Despite these relatively high conict rates, our experi-
ments show that the shared-state Omega architecture can
support service schedulers that take several seconds to make
a decision. We also investigate scaling the per-task decision
time, and found that we can support t
task
(service) of 1 second
(at a t
job
(service) of 0.1s), resulting in a conict fraction
0.2. This means that we can support schedulers with a high
one-off per-job decision time, and ones with a large per-task
decision time.
Load-balancing the batch scheduler. With the mono-
lithic single-path scheduler (4.1), the high batch job ar-
rival rate requires the use of basic, simple scheduling algo-
rithms: it simply is not possible to use smarter, more time-
consuming scheduling algorithms for these jobs as we al-
ready miss the SLO on cluster B due to the high load. Batch
jobs want to survive failures, too, and the placement quality
would doubtless improve if a scheduler could be given a little
more time to make a decision. Fortunately, the Omega archi-
360
10ms 0.1s 1s 10s 100s
10ms
1s
1m
1h
1d
Batch, avg.
Batch, 90%ile
Service, avg.
Service, 90%ile
t (service): 5ms
task
J
o
b

w
a
i
t

t
i
m
e

[
l
o
g



]
1
0
job t (service) [sec; log ]
10
(a) Job wait time.
10ms 0.1s 1s 10s 100s
0.0
1.1
2.2
3.3
4.4
5.5
Batch Service
job
t (service) [sec; log ]
10
M
e
a
n

c
o
n
f
l
i
c
t

f
r
a
c
t
i
o
n
t (service) : 5ms
task
(b) Conict fraction.
10ms 0.1s 1s 10s 100s
0.0
0.2
0.4
0.6
0.8
1.0
Batch Service no conflicts
job t (service) [sec; log ]
10
S
c
h
e
d
u
l
e
r

b
u
s
y
n
e
s
s
t (service): 5ms
task
(c) Scheduler busyness.
Figure 12: Shared-state scheduling (Omega): performance effects of varying t
job
(service) on a 7-day trace from cluster B.
10ms 0.1s 1s 10s 100s
0.0
0.2
0.4
0.6
0.8
1.0
Batch 0
single
Batch sched.
(approx.)
Batch 1
Batch 2
Service
S
c
h
e
d
u
l
e
r

b
u
s
y
n
e
s
s
job t (batch) [sec; log ]
10
t (batch): 5ms
task
(a) Scheduler busyness.
10ms 0.1s 1s 10s 100s
10ms
1s
1m
1h
1d
mean
Batch 0
90%ile
mean
Batch 1
90%ile
mean
Batch 2
90%ile
mean
Service
90%ile
job t (batch) [sec; log ]
10
J
o
b

w
a
i
t

t
i
m
e

[
l
o
g



]
1
0
t (batch): 5ms
task
(b) Job wait time.
Figure 13: Shared-state scheduling (Omega): performance
effects of splitting the batch workload across 3 batch sched-
ulers, varying t
job
(batch) in a 24h trace from cluster C.
tecture can easily achieve this by load-balancing the schedul-
ing of batch jobs across multiple batch schedulers.
To test this, we run an experiment with three parallel
batch schedulers, partitioning the workload across them by
hashing the job identiers, akin to the earlier experiment
with the simple simulator. We achieve an increase in scal-
ability of 3, moving the saturation point fromt
job
(batch)
of about 4s to 15s (Figure 13a). At the same time, the con-
ict rate remains low (around 0.1), and all schedulers meet
the 30s job wait time SLO until the saturation point (Figure
13b).
In short, load-balancing across multiple schedulers can
increase scalability to increasing job arrival rates. Of course,
the scale-up must be sub-linear due to of the overhead of
maintaining and updating the local copies of cell state, and
this approach will not easily handle hundreds of sched-
ulers. Our comparison point, however, is a single monolithic
scheduler, so even a single-digit speedup is helpful.
In summary, the Omega architecture scales well, and tol-
erates large decision times on real cluster workloads.
5.2 Dealing with conicts
We also use the high-delity simulator to explore two imple-
mentation choices we were considering for Omega.
In the rst, coarse-grained conict detection, a sched-
ulers placement choice would be rejected if any changes had
been made to the target machine since the local copy of cell
state was synchronized at the beginning of the transaction.
This can be implemented with a simple sequence number in
the machines state object.
In the second, all-or-nothing scheduling, an entire cell
state transaction would be rejected if it would cause any
machine to be over-committed. The goal here was to support
jobs that require gang scheduling, or that cannot perform any
useful work until all their tasks are running.
7
Not surprisingly, both alternatives lead to additional con-
icts and higher scheduler busyness (Figure 14). While turn-
ing on all-or-nothing scheduling for all jobs only leads to
a minor increase in scheduler busyness when using ne-
grained conict detection (Figure 14a), it does increase con-
ict fraction by about 2 as retries now must re-place all
tasks, increasing their chance of failing again (Figure 14a).
Thus, this option should only be used on a job-level gran-
ularity. Relying on coarse-grained conict detection makes
things even worse: spurious conicts lead to increases in
conict rate, and consequently scheduler busyness, by 23.
Clearly, incremental transactions should be the default.
6. Flexibility: a MapReduce scheduler
Finally, we explore how well we can meet two additional de-
sign goals of the Omega shared-state model: supporting spe-
7
This is supported by Googles current scheduler, but it is only rarely used
due to the expectation of machine failures, which disrupt jobs anyway.
361
1s 10s 100s
t
job
(service)
0.0
1.0
2.0
3.0
4.0
5.0
C
o
n

i
c
t
f
r
a
c
t
i
o
n
Coarse/Gang
Coarse/Incr.
Fine/Gang
Fine/Incr.
(a) Conict fraction.
1s 10s 100s
t
job
(service)
0.0
0.2
0.4
0.6
0.8
1.0
S
c
h
e
d
u
l
e
r
b
u
s
y
n
e
s
s
Coarse/Gang
Coarse/Incr.
Fine/Gang
Fine/Incr.
(b) Scheduler busyness.
Figure 14: Shared-state scheduling (Omega): effect of gang
scheduling and coarse-grained conict detection as a func-
tion of t
job
(service) (cluster C, 29 days); mean daily values.
cialized schedulers, and broadening the kinds of decisions
that schedulers can perform compared to the two-level ap-
proach. This is somewhat challenging to evaluate quantita-
tively, so we proceed by way of a case study that adds a
specialized scheduler for MapReduce jobs.
Cluster users at Google currently specify the number of
workers for a MapReduce job and their resource require-
ments at job submission time, and the MapReduce frame-
work schedules map and reduce activities
8
onto these work-
ers. Because the available resources vary over time and be-
tween clusters, most users pick the number of workers based
on a combination of intuition, trial-and-error and experience:
data from a months worth of MapReduce jobs run at Google
showed that frequently observed values were 5, 11, 200 and
1,000 workers.
What if the number of workers could be chosen auto-
matically if additional resources were available, so that jobs
could complete sooner? Our specialized MapReduce sched-
uler does just this by opportunistically using idle cluster re-
sources to speed up MapReduce jobs. It observes the over-
all resource utilization in the cluster, predicts the benets of
scaling up current and pending MapReduce jobs, and appor-
tions some fraction of the unused resources across those jobs
according to some policy.
MapReduce jobs are particularly well-suited to this ap-
proach because it is possible to build reasonably accurate
models of how a jobs resource allocation affects its running
time [12, 26]. About 20% of jobs in Google are MapRe-
duce ones, and many of them are run repeatedly, so historical
data is available to build models. Many of the jobs are low-
priority, best effort computations that have to make way
for higher-priority service jobs, and so may benet from ex-
ploiting spare resources in the meantime [3].
8
These are typically called tasks in literature, but we have renamed them
to avoid confusion with the cluster-scheduler level tasks that substantiate
MapReduce workers.
6.1 Implementation
Since our goal is to investigate scheduler exibility rather
than demonstrate accurate MapReduce modelling, we de-
liberately use a simple performance model that only relies
on historical data about the jobs average map and reduce
activity duration. It assumes that adding more workers re-
sults in an idealized linear speedup (modulo dependencies
between mappers and reducers), up to the point where map
activities and all reduce activities respectively run in paral-
lel. Since large MapReduce jobs typically have many more
of these activities than congured workers, we usually run
out of available resources before this point.
We consider three different policies for adding resources:
max-parallelism, which keeps on adding workers as long as
benet is obtained, global cap, which stops the MapReduce
scheduler using idle resources if the total cluster utilization
is above a target value, and relative job size, which limits the
maximum number of workers to four times as many as it ini-
tially requested. In each case, a set of resource allocations to
be investigated is run through the predictive model, and the
allocation leading to the earliest possible nish time is used.
More elaborate approaches and objective functions, such as
used in deadline-based schedulering [10], are certainly pos-
sible, but not the focus of this case study.
6.2 Evaluation
We evaluate the three different resource-allocation policies
using traces from clusters A and C, plus cluster D, which is
a small, lightly-loaded cluster that is about a quarter of the
size of cluster C. Our results suggest that 5070%of MapRe-
duce jobs can benet from acceleration using opportunistic
resources (Figure 15). The huge speedups seen in the tail
should be taken with a pinch of salt due to our simple linear
speedup model, but we have more condence in the values
for the 80
th
percentile, and here, our simulations predict a
speedup of 34 using the eager max-parallelism policy.
Although the max-parallelism policy produces the largest
improvements, the relative job size policy also does quite
well, and its speedups probably have a higher likelihood
of being achieved because it requires fewer new MapRe-
duce workers to be constructed: the time to set up a worker
on a new machine is not fully accounted for in the simple
model. The global cap policy performs almost as well as
max-parallelism in the small, under-utilized cluster D, but
achieves little or no benet elsewhere, since the cluster uti-
lization is usually above the threshold, which was set at 60%.
Adding resources to a MapReduce job will cause the
clusters resource utilization to increase, and should result
in the job completing sooner, at which point all of the jobs
resources will free up. An effect of this is an increase in the
variability of the clusters resource utilization (Figure 16).
To do its work, the MapReduce scheduler relies on being
able to see the entire clusters state, which is straightforward
in the Omega architecture. A similar argument can be made
362
0.0
0.2
0.4
0.6
0.8
1.0
max-parallel
rel. job size
global-cap
.
Job completion speedup
1x 10x 100x 0.1x
A
0.0
0.2
0.4
0.6
0.8
1.0
max-parallel
rel. job size
global-cap
Job completion speedup
1x 10x 100x 0.1x
C
00
0.2
0.4
0.6
0.8
1.0
max-parallel
rel. job size
global-cap
.
D
Job completion speedup
1x 10x 100x 0.1x
Figure 15: CDF of potential per-job speedups using different policies on clusters A, C and D (a small, lightly-utilized cluster).
CPU
RAM
0 2 4 6 8 10 12 14 16 18 20 22 24
Experiment time [hours]
U
t
i
l
i
z
a
t
i
o
n
normal
max-parallel
Figure 16: Time series of normalized cluster utilization on
cluster C without the specialized Omega MapReduce sched-
uler (top), and in max-parallelism mode (bottom).
for a specialized service scheduler for highly-constrained,
high-priority jobs. Scheduling them requires determining
which machines are applicable, and deciding how best to
place the new job while minimizing the number of preemp-
tions caused to lower-priority jobs. The shared-state model
is ideally suited to this. Our prototype MapReduce sched-
uler demonstrates that adding a specialized functionality to
the Omega system is straightforward (unlike with our cur-
rent production scheduler).
7. Additional related work
Large-scale cluster resource scheduling is not a novel chal-
lenge. Many researchers have considered this problem be-
fore, and different solutions have been proposed in the HPC,
middleware and cloud communities. We discussed several
examples in 3, and further discussed the relative merits of
these approaches in 4.
The Omega approach builds on many prior ideas. Schedul-
ing using shared state is an example of optimistic concur-
rency control, which has been explored by the database
community for a long time [18], and, more recently, consid-
ered for general memory access in the transactional memory
community [2].
Exposing the entire cluster state to each scheduler is
not unlike the Exokernel approach of removing abstractions
and exposing maximal information to applications [9]. The
programming language and OS communities have recently
revisited application level scheduling as an alternative to
general-purpose thread and process schedulers, arguing that
a single, global OS scheduler is neither scalable, nor exible
enough for modern multi-core applications demands [22].
Amoeba [3] implements opportunistic allocation of spare
resources to jobs, with motivation similar to our MapReduce
scheduler use-case. However, it achieves this by complex
communication between resource and application managers,
whereas Omega naturally lends itself to such designs as it
exposes the entire cluster state to all schedulers.
8. Conclusions and future work
This investigation is part of a wider effort to build Omega,
Googles next-generation cluster management platform.
Here, we specically focused on a cluster scheduling ar-
chitecture that uses parallelism, shared state, and optimistic
concurrency control. Our performance evaluation of the
Omega model using both lightweight simulations with syn-
thetic workloads, and high-delity, trace-based simulations
of production workloads at Google, shows that optimistic
concurrency over shared state is a viable, attractive approach
to cluster scheduling.
Although this approach will do strictly more work than a
pessimistic locking scheme as work may need to be re-done,
we found the overhead to be acceptable at reasonable operat-
ing points, and the resulting benets in eliminating head-of-
line blocking and better scalability to often outweigh it. We
also found that Omegas approach offers an attractive plat-
form for development of specialized schedulers, and illus-
trated its exibility by adding a MapReduce scheduler with
opportunistic resource adjustment.
Future work could usefully focus on ways to provide
global guarantees (fairness, starvation avoidance, etc.) in
the Omega model: this is an area where centralized control
makes life easier. Furthermore, we believe there are some
techniques from the database community that could be ap-
plied to reduce the likelihood and effects of interference for
schedulers with long decision times. We hope to explore
some of these in the future.
363
Acknowledgements
Many people contributed to the work described in this pa-
per. Members of the Omega team at Google who contributed
to this design include Brian Grant, David Oppenheimer, Ja-
son Hickey, Jutta Degener, Rune Dahl, Todd Wang and Wal-
fredo Cirne. We would like to thank the Mesos team at UC
Berkeley for many fruitful and interesting discussions about
Mesos, and Joseph Hellerstein for his early work on model-
ing scheduler interference in Omega. Derek Murray, Steven
Hand and Alexey Tumanov provided valuable feedback on
draft versions of this paper. The nal version was much im-
proved by comments from the anonymous reviewers.
References
[1] ADAPTIVE COMPUTING ENTERPRISES INC. Maui Sched-
uler Administrators Guide, 3.2 ed. Provo, UT, 2011.
[2] ADL-TABATABAI, A.-R., LEWIS, B. T., MENON, V., MUR-
PHY, B. R., SAHA, B., AND SHPEISMAN, T. Compiler and
runtime support for efcient software transactional memory.
In Proceedings of PLDI (2006), pp. 2637.
[3] ANANTHANARAYANAN, G., DOUGLAS, C., RAMAKRISH-
NAN, R., RAO, S., AND STOICA, I. True elasticity in multi-
tenant data-intensive compute clusters. In Proceedings of
SoCC (2012), p. 24.
[4] APACHE. Hadoop On Demand. http://goo.gl/px8Yd,
2007. Accessed 20/06/2012.
[5] CHANG, F., DEAN, J., GHEMAWAT, S., HSIEH, W. C.,
WALLACH, D. A., BURROWS, M., CHANDRA, T., FIKES,
A., AND GRUBER, R. E. Bigtable: A Distributed Storage
System for Structured Data. ACM Transactions on Computer
Systems 26, 2 (June 2008), 4:14:26.
[6] CHEN, Y., ALSPAUGH, S., BORTHAKUR, D., AND KATZ, R.
Energy efciency for large-scale MapReduce workloads with
signicant interactive analysis. In Proceedings of EuroSys
(2012).
[7] CHEN, Y., GANAPATHI, A. S., GRIFFITH, R., AND KATZ,
R. H. Design insights for MapReduce from diverse produc-
tion workloads. Tech. Rep. UCB/EECS201217, UC Berke-
ley, Jan. 2012.
[8] DEAN, J., AND GHEMAWAT, S. MapReduce: Simplied data
processing on large clusters. CACM 51, 1 (2008), 107113.
[9] ENGLER, D. R., KAASHOEK, M. F., AND OTOOLE, JR., J.
Exokernel: an operating system architecture for application-
level resource management. In Proceedings of SOSP (1995),
pp. 251266.
[10] FERGUSON, A. D., BODIK, P., KANDULA, S., BOUTIN, E.,
AND FONSECA, R. Jockey: guaranteed job latency in data
parallel clusters. In Proceedings of EuroSys (2012), pp. 99
112.
[11] GHODSI, A., ZAHARIA, M., HINDMAN, B., KONWINSKI,
A., SHENKER, S., AND STOICA, I. Dominant resource fair-
ness: fair allocation of multiple resource types. In Proceedings
of NSDI (2011), pp. 323336.
[12] HERODOTOU, H., DONG, F., AND BABU, S. No one (cluster)
size ts all: automatic cluster sizing for data-intensive analyt-
ics. In Proceedings of SoCC (2011).
[13] HINDMAN, B., KONWINSKI, A., ZAHARIA, M., GHODSI,
A., JOSEPH, A., KATZ, R., SHENKER, S., AND STOICA, I.
Mesos: a platform for ne-grained resource sharing in the data
center. In Proceedings of NSDI (2011).
[14] IQBAL, S., GUPTA, R., AND FANG, Y.-C. Planning con-
siderations for job scheduling in HPC clusters. Dell Power
Solutions (Feb. 2005).
[15] ISARD, M., PRABHAKARAN, V., CURREY, J., WIEDER, U.,
TALWAR, K., AND GOLDBERG, A. Quincy: fair scheduling
for distributed computing clusters. In Proceedings of SOSP
(2009).
[16] JACKSON, D. AND SNELL, Q. AND CLEMENT, M. Core al-
gorithms of the Maui scheduler. In Job Scheduling Strategies
for Parallel Processing. 2001, pp. 87102.
[17] KAVULYA, S., TAN, J., GANDHI, R., AND NARASIMHAN, P.
An analysis of traces from a production MapReduce cluster.
In Proceedings of CCGrid (2010), pp. 94103.
[18] KUNG, H. T., AND ROBINSON, J. T. On optimistic methods
for concurrency control. ACM Transactions on Database
Systems 6, 2 (June 1981), 213226.
[19] MALEWICZ, G., AUSTERN, M., BIK, A., DEHNERT, J.,
HORN, I., LEISER, N., AND CZAJKOWSKI, G. Pregel: a
system for large-scale graph processing. In Proceedings of
SIGMOD (2010), pp. 135146.
[20] MISHRA, A. K., HELLERSTEIN, J. L., CIRNE, W., AND
DAS, C. R. Towards characterizing cloud backend workloads:
insights from Google compute clusters. SIGMETRICS Perfor-
mance Evaluation Review 37 (Mar. 2010), 3441.
[21] MURTHY, A. C., DOUGLAS, C., KONAR, M., OMALLEY,
O., RADIA, S., AGARWAL, S., AND K V, V. Architecture
of next generation Apache Hadoop MapReduce framework.
Tech. rep., Apache Hadoop, 2011.
[22] PAN, H., HINDMAN, B., AND ASANOVI C, K. Lithe: en-
abling efcient composition of parallel libraries. In Proceed-
ings of HotPar (2009).
[23] PENG, D., AND DABEK, F. Large-scale incremental process-
ing using distributed transactions and notications. In Pro-
ceedings of OSDI (2010).
[24] REISS, C., TUMANOV, A., GANGER, G. R., KATZ, R. H.,
AND KOZUCH, M. A. Heterogeneity and dynamicity of
clouds at scale: Google trace analysis. In Proceedings of
SoCC (2012).
[25] SHARMA, B., CHUDNOVSKY, V., HELLERSTEIN, J., RI-
FAAT, R., AND DAS, C. Modeling and synthesizing task
placement constraints in Google compute clusters. In Pro-
ceedings of SoCC (2011).
[26] VERMA, A., CHERKASOVA, L., AND CAMPBELL, R. SLO-
driven right-sizing and resource provisioning of MapReduce
jobs. In Proceedings of LADIS (2011).
[27] WILKES, J. More Google cluster data. Google research blog,
Nov. 2011. Posted at http://goo.gl/9B7PA.
[28] ZAHARIA, M., BORTHAKUR, D., SEN SARMA, J.,
ELMELEEGY, K., SHENKER, S., AND STOICA, I. Delay
scheduling: Asimple technique for achieving locality and fair-
ness in cluster scheduling. In Proceedings of EuroSys (2010),
pp. 265278.
[29] ZHANG, Q., HELLERSTEIN, J., AND BOUTABA, R. Charac-
terizing task usage shapes in Googles compute clusters. In
Proceedings of LADIS (2011).
364