Skip to content

[WIP] Fix redundant nvml init by cached handles #56

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 1 commit into from
Closed

[WIP] Fix redundant nvml init by cached handles #56

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 13 additions & 6 deletions gpustat/__main__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,12 +11,16 @@
from .core import GPUStatCollection


def print_gpustat(json=False, debug=False, **kwargs):
def print_gpustat(json=False, debug=False, gpu_stat=None, **kwargs):
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It makes less sense to take gpu_stat as an argument for this function. I was thinking of a style like (i) we first create a GPUStatCollection instance here and (ii) then print outputs using it. As the same instance should be reused, we can consider running a wait loop inside this function rather than calling it multiple times.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What about making this function a generator, so we can keep the GPUStatCollection object in the closure.

'''
Display the GPU query results into standard output.
'''

try:
gpu_stats = GPUStatCollection.new_query()
if gpu_stat is None:
gpu_stat = GPUStatCollection()
else:
gpu_stat.update()
except Exception as e:
sys.stderr.write('Error on querying NVIDIA devices.'
Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The error catching workflow is complicated for me, can we just throw the raw error messages instead of the current one.

Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

For example, if an error happens during the loop, we can simply throw it out.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The error catching requires all the NVML API calls to be wrapped with a long try-catch block.

' Use --debug flag for details\n')
Expand All @@ -32,9 +36,10 @@ def print_gpustat(json=False, debug=False, **kwargs):
sys.exit(1)

if json:
gpu_stats.print_json(sys.stdout)
gpu_stat.print_json(sys.stdout)
else:
gpu_stats.print_formatted(sys.stdout, **kwargs)
gpu_stat.print_formatted(sys.stdout, **kwargs)
return gpu_stat


def main(*argv):
Expand Down Expand Up @@ -88,6 +93,8 @@ def main(*argv):
)
args = parser.parse_args(argv[1:])

cached_stat = None

if args.interval is None: # with default value
args.interval = 1.0
if args.interval > 0:
Expand All @@ -102,7 +109,7 @@ def main(*argv):
try:
query_start = time.time()
with term.location(0, 0):
print_gpustat(eol_char=term.clear_eol + '\n', **vars(args)) # noqa
cached_stat = print_gpustat(gpu_stat=cached_stat, eol_char=term.clear_eol + '\n', **vars(args)) # noqa
print(term.clear_eos, end='')
query_duration = time.time() - query_start
sleep_duration = args.interval - query_duration
Expand All @@ -111,7 +118,7 @@ def main(*argv):
except KeyboardInterrupt:
exit(0)
else:
print_gpustat(**vars(args))
print_gpustat(gpu_stat=cached_stat, **vars(args))


if __name__ == '__main__':
Expand Down
48 changes: 31 additions & 17 deletions gpustat/core.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,19 +250,38 @@ def jsonify(self):

class GPUStatCollection(object):

def __init__(self, gpu_list):
def __init__(self, gpu_list=[]):
"""The initialization argument gpu_list is remained to support
existing APIs"""
self.gpus = gpu_list

# attach additional system information
self.hostname = platform.node()
N.nvmlInit()
device_count = N.nvmlDeviceGetCount()
self.handles = [N.nvmlDeviceGetHandleByIndex(idx)
for idx in range(device_count)]

self.update()

def __del__(self):
# sometimes delayed gc causes problem, just attempt to release
# NVML resources
try:
N.nvmlShutdown()
except Exception:
pass

def update(self):
self._update_host()
self._update_gpu()

def _update_host(self):
"""Update additional host information"""
self.query_time = datetime.now()
self.hostname = platform.node()

@staticmethod
def new_query():
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We should not remove this static method this is already being used (as a public API, sort of) in other applications. Since you are lengthening the lifecycle of GPUStatCollection and making it update-able, we can think of supporting both constructor and factory method.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought the wrapper function gpustat.new_query() was the real API for users, instead of the implementation itself.

It's ok to retain the legacy API, but I prefer a factory function like

@classmethod
def new_query(cls):
    return cls()

def _update_gpu(self):
"""Query the information of all the GPUs on local machine"""

N.nvmlInit()

def get_gpu_info(handle):
"""Get one GPU information specified by nvml handle"""

Expand Down Expand Up @@ -365,17 +384,12 @@ def _decode(b):
return gpu_info

# 1. get the list of gpu and status
gpu_list = []
device_count = N.nvmlDeviceGetCount()
self.gpus = []

for index in range(device_count):
handle = N.nvmlDeviceGetHandleByIndex(index)
for handle in self.handles:
gpu_info = get_gpu_info(handle)
gpu_stat = GPUStat(gpu_info)
gpu_list.append(gpu_stat)

N.nvmlShutdown()
return GPUStatCollection(gpu_list)
self.gpus.append(gpu_stat)

def __len__(self):
return len(self.gpus)
Expand Down Expand Up @@ -471,7 +485,7 @@ def date_handler(obj):

def new_query():
'''
Obtain a new GPUStatCollection instance by querying nvidia-smi
Obtain a new GPUStatCollection instance by querying nvml
to get the list of GPUs and running process information.
'''
return GPUStatCollection.new_query()
return GPUStatCollection()