By extracting both monitor thread and watchdog thread into a separate class this will help us learn what dependencies we have for each thread and it will kind of simplify the consolidation work for each thread (consolidating from thread per PG instance to per PG class)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155831
Approved by: https://github.com/d4l3k, https://github.com/kwen2501
This is the start of a series of efforts to consolidating auxiliary threads in PGNCCL, aka watchdog and heartbeat_monitoring threads. Right now we launch these two threads per PG instances, i.e., if users create hundred or thousand instances of PG or subPGs, we will end up with that twice many side threads which is not efficient. We have a RFC to consolidate them (https://github.com/pytorch/pytorch/issues/146956). Right now both threads are assigned with so many functionalities so it is hard to do the consolidations in one shot, we will try to split it into at least two steps (PRs) to make it easier to test and review.
We did our first attemp in https://github.com/pytorch/pytorch/pull/153668 but we also want to try to see if we can make monitoring thread a class. This PR is doing the first step to make monitoring thread a class. The next step to also extract watchdog to be a separate class so that we know its dependency.
What we did in this PR:
1. Move all related variables and methods into a class named `HeartbeatMonitor`.
2. Correct some errors in the original logics inside monitoring thread loop.
3. Move the error propagation check to watchdog thread which is more relevant. This is totally fine since we rolled out EventCache out fully so watchdog hang is rare now.
Today there are two major functions inside heartbeat monitoring thread today:
1. Check the heartbeat of watchdog thread every 8 minutes. If no heartbeat detected and we are sure monitoring thread has not been stopped, we will kill the program by SIG_ABORT.
2. We check TCPStore every 30 sec to see if any watchdog timeout happens on other ranks, if so we will initiate a dump signal on the current rank as well. (We do this only in the default PG)
Differential Revision: [D75799278](https://our.internmc.facebook.com/intern/diff/D75799278)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153977
Approved by: https://github.com/kwen2501, https://github.com/d4l3k
This is the start of a series of efforts to consolidating auxiliary threads in PGNCCL, aka watchdog and heartbeat_monitoring threads. Right now we launch these two threads per PG instances, i.e., if users create hundred or thousand instances of PG or subPGs, we will end up with that twice many side threads which is not efficient. We have a RFC to consolidate them (https://github.com/pytorch/pytorch/issues/146956). Right now both threads are assigned with so many functionalities so it is hard to do the consolidations in one shot, we will try to split it into at least two steps (PRs) to make it easier to test and review.
We did our first attemp in https://github.com/pytorch/pytorch/pull/153668 but we also want to try to see if we can make monitoring thread a class. This PR is doing the first step to make monitoring thread a class. The next step to also extract watchdog to be a separate class so that we know its dependency.
What we did in this PR:
1. Move all related variables and methods into a class named `HeartbeatMonitor`.
2. Correct some errors in the original logics inside monitoring thread loop.
3. Move the error propagation check to watchdog thread which is more relevant. This is totally fine since we rolled out EventCache out fully so watchdog hang is rare now.
Today there are two major functions inside heartbeat monitoring thread today:
1. Check the heartbeat of watchdog thread every 8 minutes. If no heartbeat detected and we are sure monitoring thread has not been stopped, we will kill the program by SIG_ABORT.
2. We check TCPStore every 30 sec to see if any watchdog timeout happens on other ranks, if so we will initiate a dump signal on the current rank as well. (We do this only in the default PG)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153977
Approved by: https://github.com/kwen2501, https://github.com/d4l3k
Relanding #148590 due to merge conflict.
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Squashed contents:
* [ptd][nccl] use current-stream as nccl-stream under async=False mode (#147820)
PTD current workflow:
- PTD creates its own dedicated `ncclStream` for comm operation
- it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective
such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us).
This diff:
- async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead
- async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready
- pass down async from c10d down to NCCL-PG
this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%**
* [PGNCCL] Make avoid-record-stream default
* [c10d] Add asyncOp argument to Ops
* Change python side wait
* Pass asyncOp at ProcessGroup level
* Watchdog unstashing tensors as a safety net
* Stash tensors for reduce_scatter_v and all_gather_v
Pull Request approved: https://github.com/pytorch/pytorch/pull/149753
* [c10d] Move unstashing from watchdog to main thread
Pull Request approved: https://github.com/pytorch/pytorch/pull/150079
* [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation
Pull Request approved: https://github.com/pytorch/pytorch/pull/150130
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150398
Approved by: https://github.com/atalman
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70937982](https://our.internmc.facebook.com/intern/diff/D70937982)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70835197](https://our.internmc.facebook.com/intern/diff/D70835197)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70835197](https://our.internmc.facebook.com/intern/diff/D70835197)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
Summary:
This PR is basically a replacement of
https://github.com/pytorch/pytorch/pull/140087, which caused some perf
drop due to frequent TCPStore check in watchdog thread. The fix is to move the
tcpstore check in monitoring thread
If unhealthy, the user should be able to get the type of errors, e.g.,
timeout,nccl error or remote error.
This API is applied to PG level, compared to the
work.get_future_result() API which is applied to Work Level.
Error detection at PG level is much more convenient for users to handle
the PG failure as a whole, e.g, restarting the PG.
Error handling at the work level is still useful for users to attach
work specific context and debug the RC of the specific failing
work/collective
Note it is critical for all ranks in the PG to be notified about an
error as soon as it occurs, so we introduce an errorType of
REMOTE_ERROR, which is 'broadcasted' from a src rank (which detects a
local error) to all other ranks in the PG, the broadcast is done through
TCPStore currently
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144498
Approved by: https://github.com/kwen2501
- Refactored traceback code into `work.printTraceback()`. cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @shuqiangzhang
- Refactored desync debug code into `class DesyncDebugger`.
- Moved occurrences of `futureWorkResult_->markCompleted` into `checkAndSetException` and `checkTimeout`, respectively. cc @shuqiangzhang
- Modularized dump signal broadcast code into `ProcessGroupNCCL::broadcastDumpSignal`. cc @fduwjj @c-p-i-o
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139834
Approved by: https://github.com/shuqiangzhang
Summary:
Blocking wait mode is not widely used, probably useful in debugging.
in blockingWait mode, we don't need to enable the watchdog thread to
check the timeout or nccl error because the main thread would throw an
exception if error happens and it is obvious to user which work fails
and its user's responsibility to handle the exception.
Test Plan:
CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138001
Approved by: https://github.com/fduwjj, https://github.com/c-p-i-o
ghstack dependencies: #137799
synchronization
Summary:
Barrier is essentially intended to block CPU thread (instead of GPU
streams). Before we used 2 stream synchronizations (1. current stream
blocked by nccl stream end event, 2. CPU thread blocked on current
stream). This is unnecessary as we already have CPU thread blocking
logic in wait(). Also, adding barrier specific code block in the general
GPU synchronize() API is intrusive and confusing.
This PR cleans this.
Test Plan:
CI
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/137516
Approved by: https://github.com/fduwjj, https://github.com/kwen2501
This adds logs if we can't acquire locks in NCCLUtils and ProcessGroupNCCL for 30s.
This is motivated by some deadlocks were seeing and it's unclear if it's in NCCL or on the PyTorch side of things.
This required replacing most `std::mutex` with `std::timed_mutex` and `std::condition_variable_any` as appropriate.
Test plan:
existing CI for regressions
will add unit tests on `C10D_LOCK_GUARD`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134131
Approved by: https://github.com/c-p-i-o, https://github.com/fduwjj
Summary: Pass process group info into NcclWork
Test Plan: buck2 run mode/dev-nosan kineto/libkineto/fb/integration_tests:pytorch_execution_trace_integration_test
Differential Revision: D61677160
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134269
Approved by: https://github.com/wconstab
This adds logs if we can't acquire locks in NCCLUtils and ProcessGroupNCCL for 30s.
This is motivated by some deadlocks were seeing and it's unclear if it's in NCCL or on the PyTorch side of things.
This required replacing most `std::mutex` with `std::timed_mutex` and `std::condition_variable_any` as appropriate.
Test plan:
existing CI for regressions
will add unit tests on `C10D_LOCK_GUARD`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134131
Approved by: https://github.com/c-p-i-o, https://github.com/fduwjj
Summary:
Split out `seq_id` into `collective_seq_id` and `p2p_seq_id`. The main idea here is that collectives that go to all machines should have identical `collective_seq_id` and therefore it makes it easier to spot if one of machines isn't handling a collective operation.
Next, we can attempt to match up p2p operations to ensure that the sender(s)/receivers(s) are in sync.
Resolves issue: https://github.com/pytorch/pytorch/issues/125173
Test Plan:
Unit tests.
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125727
Approved by: https://github.com/zdevito
Summary:
We found that some dumps are missing when monitoring thread timeout.
This is likely due to multiple PGs could still dump the same records
at the same time. So we should allow only PG0 to actualy dump
Test Plan:
unit test
python test/run_test.py --cpp --verbose -i cpp/ProcessGroupNCCLErrorsTest
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125356
Approved by: https://github.com/c-p-i-o
Fix and test issues with both coalesced and individual send/recv ops
Considered an alternate approach and then ditched it
- alternate approach: #119757
- reason ditched: prefer recording individual collective events inside
coalescing region instead of just the event at the end of the region,
which also would not have tensor sizes or opnames without additional
state variables added
Another approach also ditched
- record events on workEnqueue instead of initWork
- reason ditched: too messy to get input/output shapes tagged on
recording when recording in workEnqueue. Adding the info onto the
Work obj would be possible, but adds to overhead of copying Works
which we do on every collective. We can get info off the input/output
tensors directly in initWork, but we don't want to keep refs to those
tensors alive while the work is Enqueued, so we'd have to specifically
copy size lists or something.
This PR instead avoids creating a work inside pointToPoint when
coalescing is active. Instead, only at endCoalescing() is a work finally
intialized and enqueued. But it adds a record() call inside
pointToPoint() instead of creating a work, during coalescing. This
record() call picks up tensor shapes and op names.
It ALSO changes initWork to accept a 'record' argument. This defaults to
false, and should only be set to true if the caller ensures the work
will be enqueued by workEnqueue, ensuring its cuda events are live when
used by flight recorder's update_state().
The testing uncovers some odd pre-existing behavior and leaves them
alone for now. We could change some of these
- seq starts off at 1, not 0 for first op (but this is inconistent)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120270
Approved by: https://github.com/shuqiangzhang
ghstack dependencies: #120724
Summary:
The current dump timeout logic is a bit cumbersome as it needs 2 times: 1.
timeout, 2. wake up time. And in theory the caller just needs to wait
for a max of timeout value for the dump and declare the dump to be
either successful or not. Also we unify the async call using std::async
instead of a customized async lauch function for each operation.
Test Plan:
Unit tests
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120331
Approved by: https://github.com/wconstab
Today watchdog's sleep interval is 1s. That's a bit long compared to modern GPU link's (or network link's) speed.
Take DDP and Ampere for example:
DDP's bucket size = 25 MB
Ampere's NVLink speed = 250 GB/s
25 MB / 250 GB/s = 100 ms.
So we are updating the interval to 100 ms.
Update:
25 MB / 250 GB/s = 0.1 ms
But let's see how it goes so far between making the checking more aggressive.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117297
Approved by: https://github.com/fduwjj
Previously, we have the writer register to each NCCL PG(backend), so for every pg, we have a NCCL PG instance, so if we use some customized writer when multiple sub-PGs are used, we need to ensure user to register the writer for every backend which indicates a bad UX. Furthermore, the debug info is global, so it does not make sense to have the writer for each instance. We even have a static mutex in the `dumpDebuggingInfo` to ensure we serialize the write, that makes it more obvious that we can make the writer a singleton so that we only have one writer instance for all PG instances.
Although the rationale is clear, the implementation may vary a lot. So this PR is RFC for now to see if this implementation makes sense or not.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116489
Approved by: https://github.com/kwen2501
Replaces the "always sleep 30 sec before abort" with "wait up to 30 sec
for the future to complete then abort". The difference in this case is
the abort happens as soon as the dump finishes up to a maximum, instead
of always waiting the maximum.
Allows multiple calls to dump, which will be serialized.
Renames tryWriteDebugInfo to launchAsyncDebugDump in spirit of the
change to support more than one launch and to always launch rather than
only launching on the first call.
Adds a test for dumping on timeout.
This reverts commit ac7d14baad53fa7d63119418f760190f289d8a01.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115332
Approved by: https://github.com/fduwjj
Replaces the "always sleep 30 sec before abort" with "wait up to 30 sec
for the future to complete then abort". The difference in this case is
the abort happens as soon as the dump finishes up to a maximum, instead
of always waiting the maximum.
Allows multiple calls to dump, which will be serialized.
Renames `tryWriteDebugInfo` to `launchAsyncDebugDump` in spirit of the
change to support more than one launch and to always launch rather than
only launching on the first call.
Adds a test for dumping on timeout.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115176
Approved by: https://github.com/zdevito
NCCL_ prefix should only be used for NCCL library's environment variables. We currently use a few environment variables in PyTorch with the NCCL_ prefix that are the NCCL library does not understand.
This patch renames such environment variables to use the TORCH_NCCL_ prefix instead. We still maintain the old NCCL_ variables, but throw a warning when they are used.
The following env changes have been made:
`NCCL_BLOCKING_WAIT` -> `TORCH_NCCL_BLOCKING_WAIT`
`NCCL_ENABLE_TIMING` -> `TORCH_NCCL_ENABLE_TIMING`
`NCCL_DESYNC_DEBUG` -> `TORCH_NCCL_DESYNC_DEBUG`
`NCCL_ASYNC_ERROR_HANDLING` -> `TORCH_NCCL_ASYNC_ERROR_HANDLING`
`ENABLE_NCCL_HEALTH_CHECK` -> `TORCH_ENABLE_NCCL_HEALTH_CHECK`
`NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK` -> `TORCH_NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK`
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114077
Approved by: https://github.com/fduwjj
Summary:
The getCvar* functions allow us to provide multiple environment variables for the same value. This allows us to deprecate some variables in favor of others, while still allowing users to temporarily use the old variables for some time.
Test Plan: OSS CI
Reviewed By: fduwjj, XilunWu
Differential Revision: D51225487
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113797
Approved by: https://github.com/fduwjj
This PR is to enable the store of NCCL flight recorder to storage and make it configurable by letting users register their own way of storing the debug info. We will then provide users a script to offline parse and process the dumped blobs.
One thing, this PR is not trying to resolve is to decide where to dump the debug info. I will send a follow-up PR to address that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113503
Approved by: https://github.com/zdevito
This PR has the following goals:
1. Detect unhealthy nccl watchdog thread by implementing a heartbeat. NCCL watchdog sometimes can hang for several reasons such as nccl/cuda API bugs or unexpected blocking behaviors. This is the last resort to ensure that we don't silently keep the training job run for hours.
2. Sometimes, the process gets stuck in the destroy of NCCL PG, and this PR will ensure that we will eventually abort it after some time (by default 2 mins)
3. Once heartbeat cannot be heard, we dump debug information (for now, we just use the flight recorder implemented in https://github.com/pytorch/pytorch/pull/110960/files) to disk. (How and where to dump the debug info will be addressed in the following PR).
4. Finally, we initiate std::abort via `LOG(FATAL)` to kill the process.
To clarify further what this PR is trying to solve, we first list are four cases when a NCCL PG can end up with:
- case 1: ncclwatchdog gets stuck (maybe some blocking API) and heartbeat monitor kills it during regular heartbeat monitor loop.
- case 2: ncclwatchdog timeout and desync report or destroy kicked in(let's call it shutdown) but this shutdown takes so long and heartbeat believes it has to kills the process anyway.
- case 3: ncclwatchdog aborts the process (heartbeat monitor not involved)
- case 4: program exits cleanly (heartbeat monitor not involved)
As we can see here, this PR is trying to address case one and two and we also want to ensure adding one more monitor thread does not interfere what we are currently doing in case three and four. That's why we added two flags `terminateHeartbeatMonitorThread_` and `collectiveDebugInfoMode_`.
For case three and four, either `monitorWakeUpCV_` will be waked up in the destructor or `terminateHeartbeatMonitorThread_` will be set to true. So that monitor thread will just exit ASAP.
For case one, both `terminateHeartbeatMonitorThread_` and `collectiveDebugInfoMode_` will still false when monitor thread see there are no heartbeat, so it will directly kill the process. For case two, either `terminateHeartbeatMonitorThread_` and `collectiveDebugInfoMode_` will be true, the monitor thread will wait extra time before killing the process.
Differential Revision: [D51146305](https://our.internmc.facebook.com/intern/diff/D51146305)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112518
Approved by: https://github.com/kwen2501, https://github.com/wconstab
Keep a buffer of the last 16384 nccl work actions, including the stack
trace that launched the event.
When torch._C._distributed_c10d._dump_nccl_trace(), it an dump these to
a pickled archive.
For each action we get:
process_group_id, seq_id, collective_name, size_of_first_tensor, stack trace
state - issued, started, completed (based on cuda events and queried if
necessary when the dump is requested)
I tested that it is possible to query event state when the streams are
otherwise stuck.
Differential Revision: [D50138956](https://our.internmc.facebook.com/intern/diff/D50138956)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110960
Approved by: https://github.com/wconstab
Fixes#97191
This PR aims to propagate collective exceptions (async error or timeout) up to the program, so as to avoid silent stuck job.
### Previous output in #97191
```
Rank 0 is the problematic rank
Rank 4 completed
Rank 5 completed
Rank 3 completed
Rank 6 completed
Rank 2 completed
Rank 7 completed
Rank 1 completed
[E ProcessGroupNCCL.cpp:464] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=1, OpType=ALLREDUCE, Timeout(ms)=10000) ran for 10917 milliseconds before timing out.
Rank 0 completed
[E ProcessGroupNCCL.cpp:478] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
[E ProcessGroupNCCL.cpp:483] To avoid data inconsistency, we are taking the entire process down.
```
Although it says that it is taking the process down, it sometimes fails to do so.
### New output after this PR:
```
...
[E ProcessGroupNCCL.cpp:459] [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=1, OpType=ALLREDUCE, Timeout(ms)=10000) ran for 10599 milliseconds before timing out.
[E ProcessGroupNCCL.cpp:473] Some NCCL operations have failed or timed out. Due to the asynchronous nature of CUDA kernels, subsequent GPU operations might run on corrupted/incomplete data.
[E ProcessGroupNCCL.cpp:479] To avoid data inconsistency, we are taking the entire process down.
[E ProcessGroupNCCL.cpp:818] [Rank 0] NCCL watchdog thread terminated with exception: [Rank 0] Watchdog caught collective operation timeout: WorkNCCL(SeqNum=1, OpType=ALLREDUCE, Timeout(ms)=10000) ran for 10599 milliseconds before timing out.
ERROR:torch.distributed.elastic.multiprocessing.api:failed (exitcode: -6) local_rank: 0 (pid: 194470) of binary: /data/home/kw2501/repos/pytorch-dev-env/bin/python
Traceback (most recent call last):
File "/pytorch-dev-env/bin/torchrun", line 33, in <module>
sys.exit(load_entry_point('torch', 'console_scripts', 'torchrun')())
File "/pytorch-dev/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
return f(*args, **kwargs)
File "/pytorch-dev/torch/distributed/run.py", line 794, in main
run(args)
File "/pytorch-dev/torch/distributed/run.py", line 785, in run
elastic_launch(
File "/pytorch-dev/torch/distributed/launcher/api.py", line 134, in __call__
return launch_agent(self._config, self._entrypoint, list(args))
File "/pytorch-dev/torch/distributed/launcher/api.py", line 250, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
hang.py FAILED
------------------------------------------------------------
Failures:
<NO_OTHER_FAILURES>
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-03-20_22:00:42
host : node0
rank : 0 (local_rank: 0)
exitcode : -6 (pid: 194470)
error_file: <N/A>
traceback : Signal 6 (SIGABRT) received by PID 194470
============================================================
```
The log suggests that TorchX monitor is triggered, and job is torn down.
### Major changes in this PR:
1. Merge ncclWatchDog thread and workCleanupLoop thread into one so that the watch action and the throw action are streamlined.
Previously, ncclWatchDog is responsible for watching comm error and timeout, and workCleanupLoop is responsible for watching Work item error and throwing exception. This two-thread design is not streamlined, raising the chance of missing the throw. Also, it is duplicated to watch at multiple level.
2. Rethrow exception at watchdog thread.
3. Clean up a bunch of duplicated functions, e.g. `checkAndThrowException` and `handleNcclException`.
4. Turn on ASYNC_ERROR_HANDLING by default
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97066
Approved by: https://github.com/rohan-varma
