Closed Sms-Rk closed 1 year ago
we needed to accept interval less than number 1 so i did very very simple change to accept this
Hi @Sms-Rk, sorry that I'm going to close this. I'm pretty sure that this change will not accomplish your need. There is also an old discussion in #32 for a similar feature.
Here is something nvitop
can do:
<Enter>
key. The metrics-watching screen will update at an interval of 1/4 sec.
Watch metrics for a specific process (shortcut: Enter / Return).
nvitop.ResourceMetricCollector
, see Resource Metric Collector for more information.If you are working on a delay-sensitive application, I strongly suggest you use a proper profiler rather than a monitor.
FYI, the true interval is defined by TTLCache
s. They always run at intervals of 1 second.
Personally, I do not prefer float intervals.
Python <-> NVML C API
, NVML C API <-> NVIDIA driver
. You always get delayed and inaccurate results. Especially when your snapshot interval is small. If the delay matters (if this is why you want this feature is this PR), you should use a profiler.we needed to accept interval less than number 1 so i did very very simple change to accept this
Hi @Sms-Rk, sorry that I'm going to close this. I'm pretty sure that this change will not accomplish your need. There is also an old discussion in #32 for a similar feature.
Here is something
nvitop
can do:
- Select a process and then press the
<Enter>
key. The metrics-watching screen will update at an interval of 1/4 sec.Watch metrics for a specific process (shortcut: Enter / Return).
- Use
nvitop.ResourceMetricCollector
, see Resource Metric Collector for more information.If you are working on a delay-sensitive application, I strongly suggest you use a proper profiler rather than a monitor.
FYI, the true interval is defined by
TTLCache
s. They always run at intervals of 1 second.Personally, I do not prefer float intervals.
- The CLI UI is updated at intervals of 1/4 second. Any interval lower than that will have no effect but will consume much more resources.
- The API calls and communication always have delays, for example,
Python <-> NVML C API
,NVML C API <-> NVIDIA driver
. You always get delayed and inaccurate results. Especially when your snapshot interval is small. If the delay matters (if this is why you want this feature is this PR), you should use a profiler.
hi @XuehaiPan thanks for your attention ok, we need to monitor and profile SM metric faster than 1 sec , maybe 0.1 sec how can we do it?
we need to monitor and profile SM metric faster than 1 sec , maybe 0.1 sec
See my comment above. For delay sensitive use cases, you should use a proper profiling tool rather than a monitor. Profilers need in-process injection. This is not a monitor can do.
Implemented in #67.
hi, we needed to accept interval less than number 1 so i did very very simple change to accept this maybe its not completely right but if it is not proper to merge, i will appriciate to do this for us