InfiniBand for the enterprise




InfiniBand for the enterprise, a customer’s perspective.
                 Eric Ulmer, Accenture



              www.openfabrics.org
Agenda


   Company Overview
   Our Platform/Environment
   Our Rewards in using the OpenFabrics
   Software
   Wish List
   Q&A (Will try and maximize alotted time for
   this portion)


www.openfabrics.org                              2
Accenture Software Utility Services

   We are an Applications Service Provider
        Example customers:
          • Navitaire
               • Reservations, Crew Scheduling, Online Bookings, Customer
                 Relationship Management, Route Solver
          • Insurance packages
          • Telecommunication company software packages
               • Mobile telephone number portability
          • Utility company management software packages
               • Natural Gas and Electric trouble ticket management
          • Custom stock market analysis and trading packages

www.openfabrics.org                                                         3
Accenture Software Utility Services




www.openfabrics.org                   4
Accenture Software Utility Services

   8 Data Centers
        2 in United States
        3 in the United Kingdom
        3 in Australia
        400+ UNIX/Linux Servers
        2000+ Windows Instances
        Heavy I/O Microsoft SQL shop (350+ instances)
        1.6PB of disk total.
        100% Network Appliance primary and nearline disk (54 filers)
          • NFS for VMWARE and UNIX databases, iSCSI and FCP for Windows
        Data Domain being used for VTL
        HP, SUN and IBM servers



www.openfabrics.org                                                        5
Navitaire
   High Transactional Volume                   $11.6B
                                               $10.7B
   Total value of web-based travel                     2006


   transactions exceeds Amazon.com transactions
   Daily credit authorizations reach $135M
   Execute 750,000 batch jobs monthly
   High Service level Agreement penalties:
        Outages
        Application response time
        We have no scheduled maintenance windows. We pay
        to take Airline customers down.


www.openfabrics.org                                           6
Our Platform(s)
   Traditional HP 5XX & 3XX Servers (Web/App/DB)
        Gig-E/Mostly iSCSI/4-16CPU/16-128GB RAM
        Windows 2003 Std and Enterprise Edition; RHEL 3/4

   IBM X460 and X3950(DB)
        Windows 2003 Enterprise Edition
        Emulex FCP/Teamed Gig-E/16-32 core/128GB RAM

   SUN Sparc (Misc)
        Gig-E and 100Mb ethernet connected. Various generations of models;
        Solaris 8-10

   SUN X86 4600M2 (VMWare)
        ConnectX - DDR; VMWare ESX 3.5U3; 32 Cores; 256GB Ram each
        Xsigo VNIC driver on top of OF driver stack.



www.openfabrics.org                                                          7
Our Platforms
   Cisco Ethernet Switching and Routing equipment
       Primarily GigE to servers; 10G for trunks and Filer connectivity
       Dual Switches; HSRP (Full redundancy)
       Servers patched to core via patch panels
   Brocade and Cisco fiber switches
       Dual fabrics; models vary by data center and generation.
   InfiniBand based LAN/SAN virtualization
       Two separate IB fabrics and Xsigo director cores per data center
       Used as our backbone for all VMWare machines
       Our VMWare farms do not have ethernet connections except for
       remote consoles; 100% InfiniBand based VMWare deployment.
       Each VMware server has one connection to each InfiniBand fabric
       via separate ConnectX HCAs.
       Xsigo provides virtual network and fiber HBAs to the vmware servers
       over InfiniBand via their VP780 director modules.
       We are using Mellanox Silicon with the OpenFabrics ESX drivers

www.openfabrics.org                                                      8
Our Platforms

   VMWare Platform




www.openfabrics.org   9
Rewards using OF components

   Cost Savings!
        Neutral to our present 1GbE deployment with DDR(20g)
        Cost is less than 1GbE if using Single Data Rate
        InfiniBand(SDR IB) (10g)
        Cost 4-12X lower than 10GbE with Quad Data Rate
        InfiniBand(QDR IB) edge(40g) and DDR(20g) Core
   Reduced deployment time!
        Two quick connections to an IB edge switch –vs- multiple
        GigE and FC patch connections. (Not to mention the time
        and planning in installing the actual patch panels)
        Can deploy “data center in a carton”
        2-4 hours vs potentially 2 months in some cases.


www.openfabrics.org                                            10
BEFORE




                               Picture not of actual
                               Navitaire patch panel


www.openfabrics.org                                    11
AFTER




www.openfabrics.org   12
Rewards using OF components

   Eye to the future
        We are using Solid State disk storage in production in
        one of our data centers today
          • 8TB usable of Texas Memory Systems RamSan 500 storage
            front-ended by NetApp V6080 filer today.
          • We foresee a future without “LUNS” on servers
               • Why access RAM (Flash or otherwise) as a disk device?
               • Traditional Ethernet is obviously not the right choice for extended or
                 RDMA memory access
          • InfiniBand and OpenFabrics are obvious existing (and more
            importantly, shipping) choices for transmitting at high data rates
            at the lowest cost.
          • Our goal is to be off spinning disk for all primary applications by
            2012.

www.openfabrics.org                                                                  13
Rewards in using OF components
   Eye to the future
        We are testing communication from the application tier to database
        server tier using Microsoft winsock direct and IPoIB from the WinOF
        driver package.
        Our database servers have large RAM caches & they can bury a
        standard ethernet network queue. This causes a troublesome square
        toothed performance curve; unpredictable response times.
        WinOF provides:
          • Lower latency
          • Higher data rates (7.2Gb/sec SDR vs our current 420Mb/sec Quad
            GieG NIC Team)
                • We’re excited about QDR; our purchases are road mapped.
          •   Lower CPU consumption (2% vs 40% with NIC Team)
          •   Lower cost
          •   Lower complexity
          •   A predictable sine wave shaped performance curve


www.openfabrics.org                                                          14
Challenges (WinOF)
   Limited Windows HCL and “designed for” logo certification
   of drivers for enterprise editions of Windows. Very HPC
   centric in orientation.
        If it’s not blessed by a Microsoft Logo it cannot be used in the
        Enterprise.
        Enterprise customers have six to seven digit support contracts with
        Microsoft which are useless w/o Microsoft blessed software.

   Installation and documentation is centered on experienced
   computer users.
        Needs to be GUIfied and dumbed down for Enterprise/End-User
        Windows admins




www.openfabrics.org                                                           15
What’s missing in the ecosphere?

   No existing mainstream enterprise storage vendor doing
   native IPoIB today.
        Lots of niche HPC offerings; nothing for the enterprise
        Should be top priority to get EMC, NetApp, SUN or HDS onboard
        with native IB connections.
   RDMA/Connected mode type connectivity is needed to
   unlock the performance potential of the solid state storage
   era.




www.openfabrics.org                                                     16
Wish List for the ecosphere
   Windows Certifications for WinOF
       Clustering catalog entries
       Base installer and documentation to be more like other windows
       software installs (Next->Next->Next)
       GUID->MAC mapping always auto-magic in WinOF
   Vendors need to make sure their products are submitted to Microsoft for
   inclusion in their catalog (WHQL)
   Enterprise Storage Vendor (we need one!)
       One vendor is on the fence now
       Contact your storage vendor and ask for InfiniBand
   Connected Mode or NFS/RDMA IO to VMWare
       Customers contact your VMWare rep and ask for InfiniBand
       connectivity
   VLAN and Private VLAN like seperation and security
       Required to separate different tiers of the stack; multiple customer
       segmentation
       Partition key management?
www.openfabrics.org                                                       17
Q&A




                      Q&A




www.openfabrics.org         18

InfiniBand for the enterprise

  • 1.
    InfiniBand for theenterprise InfiniBand for the enterprise, a customer’s perspective. Eric Ulmer, Accenture www.openfabrics.org
  • 2.
    Agenda Company Overview Our Platform/Environment Our Rewards in using the OpenFabrics Software Wish List Q&A (Will try and maximize alotted time for this portion) www.openfabrics.org 2
  • 3.
    Accenture Software UtilityServices We are an Applications Service Provider Example customers: • Navitaire • Reservations, Crew Scheduling, Online Bookings, Customer Relationship Management, Route Solver • Insurance packages • Telecommunication company software packages • Mobile telephone number portability • Utility company management software packages • Natural Gas and Electric trouble ticket management • Custom stock market analysis and trading packages www.openfabrics.org 3
  • 4.
    Accenture Software UtilityServices www.openfabrics.org 4
  • 5.
    Accenture Software UtilityServices 8 Data Centers 2 in United States 3 in the United Kingdom 3 in Australia 400+ UNIX/Linux Servers 2000+ Windows Instances Heavy I/O Microsoft SQL shop (350+ instances) 1.6PB of disk total. 100% Network Appliance primary and nearline disk (54 filers) • NFS for VMWARE and UNIX databases, iSCSI and FCP for Windows Data Domain being used for VTL HP, SUN and IBM servers www.openfabrics.org 5
  • 6.
    Navitaire High Transactional Volume $11.6B $10.7B Total value of web-based travel 2006 transactions exceeds Amazon.com transactions Daily credit authorizations reach $135M Execute 750,000 batch jobs monthly High Service level Agreement penalties: Outages Application response time We have no scheduled maintenance windows. We pay to take Airline customers down. www.openfabrics.org 6
  • 7.
    Our Platform(s) Traditional HP 5XX & 3XX Servers (Web/App/DB) Gig-E/Mostly iSCSI/4-16CPU/16-128GB RAM Windows 2003 Std and Enterprise Edition; RHEL 3/4 IBM X460 and X3950(DB) Windows 2003 Enterprise Edition Emulex FCP/Teamed Gig-E/16-32 core/128GB RAM SUN Sparc (Misc) Gig-E and 100Mb ethernet connected. Various generations of models; Solaris 8-10 SUN X86 4600M2 (VMWare) ConnectX - DDR; VMWare ESX 3.5U3; 32 Cores; 256GB Ram each Xsigo VNIC driver on top of OF driver stack. www.openfabrics.org 7
  • 8.
    Our Platforms Cisco Ethernet Switching and Routing equipment Primarily GigE to servers; 10G for trunks and Filer connectivity Dual Switches; HSRP (Full redundancy) Servers patched to core via patch panels Brocade and Cisco fiber switches Dual fabrics; models vary by data center and generation. InfiniBand based LAN/SAN virtualization Two separate IB fabrics and Xsigo director cores per data center Used as our backbone for all VMWare machines Our VMWare farms do not have ethernet connections except for remote consoles; 100% InfiniBand based VMWare deployment. Each VMware server has one connection to each InfiniBand fabric via separate ConnectX HCAs. Xsigo provides virtual network and fiber HBAs to the vmware servers over InfiniBand via their VP780 director modules. We are using Mellanox Silicon with the OpenFabrics ESX drivers www.openfabrics.org 8
  • 9.
    Our Platforms VMWare Platform www.openfabrics.org 9
  • 10.
    Rewards using OFcomponents Cost Savings! Neutral to our present 1GbE deployment with DDR(20g) Cost is less than 1GbE if using Single Data Rate InfiniBand(SDR IB) (10g) Cost 4-12X lower than 10GbE with Quad Data Rate InfiniBand(QDR IB) edge(40g) and DDR(20g) Core Reduced deployment time! Two quick connections to an IB edge switch –vs- multiple GigE and FC patch connections. (Not to mention the time and planning in installing the actual patch panels) Can deploy “data center in a carton” 2-4 hours vs potentially 2 months in some cases. www.openfabrics.org 10
  • 11.
    BEFORE Picture not of actual Navitaire patch panel www.openfabrics.org 11
  • 12.
  • 13.
    Rewards using OFcomponents Eye to the future We are using Solid State disk storage in production in one of our data centers today • 8TB usable of Texas Memory Systems RamSan 500 storage front-ended by NetApp V6080 filer today. • We foresee a future without “LUNS” on servers • Why access RAM (Flash or otherwise) as a disk device? • Traditional Ethernet is obviously not the right choice for extended or RDMA memory access • InfiniBand and OpenFabrics are obvious existing (and more importantly, shipping) choices for transmitting at high data rates at the lowest cost. • Our goal is to be off spinning disk for all primary applications by 2012. www.openfabrics.org 13
  • 14.
    Rewards in usingOF components Eye to the future We are testing communication from the application tier to database server tier using Microsoft winsock direct and IPoIB from the WinOF driver package. Our database servers have large RAM caches & they can bury a standard ethernet network queue. This causes a troublesome square toothed performance curve; unpredictable response times. WinOF provides: • Lower latency • Higher data rates (7.2Gb/sec SDR vs our current 420Mb/sec Quad GieG NIC Team) • We’re excited about QDR; our purchases are road mapped. • Lower CPU consumption (2% vs 40% with NIC Team) • Lower cost • Lower complexity • A predictable sine wave shaped performance curve www.openfabrics.org 14
  • 15.
    Challenges (WinOF) Limited Windows HCL and “designed for” logo certification of drivers for enterprise editions of Windows. Very HPC centric in orientation. If it’s not blessed by a Microsoft Logo it cannot be used in the Enterprise. Enterprise customers have six to seven digit support contracts with Microsoft which are useless w/o Microsoft blessed software. Installation and documentation is centered on experienced computer users. Needs to be GUIfied and dumbed down for Enterprise/End-User Windows admins www.openfabrics.org 15
  • 16.
    What’s missing inthe ecosphere? No existing mainstream enterprise storage vendor doing native IPoIB today. Lots of niche HPC offerings; nothing for the enterprise Should be top priority to get EMC, NetApp, SUN or HDS onboard with native IB connections. RDMA/Connected mode type connectivity is needed to unlock the performance potential of the solid state storage era. www.openfabrics.org 16
  • 17.
    Wish List forthe ecosphere Windows Certifications for WinOF Clustering catalog entries Base installer and documentation to be more like other windows software installs (Next->Next->Next) GUID->MAC mapping always auto-magic in WinOF Vendors need to make sure their products are submitted to Microsoft for inclusion in their catalog (WHQL) Enterprise Storage Vendor (we need one!) One vendor is on the fence now Contact your storage vendor and ask for InfiniBand Connected Mode or NFS/RDMA IO to VMWare Customers contact your VMWare rep and ask for InfiniBand connectivity VLAN and Private VLAN like seperation and security Required to separate different tiers of the stack; multiple customer segmentation Partition key management? www.openfabrics.org 17
  • 18.
    Q&A Q&A www.openfabrics.org 18