Process-wide channel cache for gRPC+aio #120
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
justheuristic
force-pushed
the
dht_update_nov
branch
2 times, most recently
from
e7ca5e9 to
e9e47d1
Compare
justheuristic
changed the title
justheuristic
force-pushed
the
dht_update_nov
branch
from
36efba0 to
f6b484e
Compare
mryab
requested changes
Nov 17, 2020
hivemind/client/allreduce.py
Outdated
| @@ -150,9 +150,7 @@ async def accumulate(self, source: Endpoint, part: torch.Tensor) -> torch.Tensor | |||
| return await self.averaged_part | |||
|
|
|||
| def _get(self, peer: Endpoint) -> averaging_pb2_grpc.DecentralizedAveragingStub: | |||
There was a problem hiding this comment.
I believe _get is slightly overloaded here, how about _get_peer_stub?
There was a problem hiding this comment.
its mostly a convention with DHTProtocol._get, should we rename both?
Comment on lines
+4
to
+7
| from __future__ import annotations | ||
| import os | ||
| import threading | ||
| from typing import NamedTuple, Sequence, Tuple, Optional, Union, Any, Dict, TypeVar, Type |
There was a problem hiding this comment.
If we're using from __future__ import annotations, do we need to import tuple and dict from typing?
There was a problem hiding this comment.
i agree, but i would request to keep it as is AND write a todo for #98 , ok?
Comment on lines
+39
to
+41
| MAXIMUM_CHANNELS = os.environ.get("GRPC_PYTHON_MANAGED_CHANNEL_MAXIMUM", 4096) | ||
| EVICTION_PERIOD_SECONDS = os.environ.get("GRPC_PYTHON_MANAGED_CHANNEL_EVICTION_SECONDS", 10 * 60) | ||
| logger.debug(f"Eviction period = {EVICTION_PERIOD_SECONDS}s, max channels = {MAXIMUM_CHANNELS}") |
There was a problem hiding this comment.
These constants should be module-level
There was a problem hiding this comment.
upd - it complicates testing, added to #98
hivemind/utils/grpc.py
Outdated
Comment on lines
47
to
49
| _eviction_thread: threading.Thread | ||
| _nearest_expiration_time: DHTExpiration | ||
| _is_active: bool |
There was a problem hiding this comment.
These are instance attributes and are declared during init
Co-authored-by: Max Ryabinin <[email protected]>
Co-authored-by: Max Ryabinin <[email protected]>
… into dht_update_nov
mryab
approved these changes
Nov 17, 2020
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
process-wide channel caching
Currently, each time we query a remote peer with gRPC, we have to create a new channel.
In contrast, gRPC best practices recommend reusing channels for multiple rpc calls
hivemind/client/expert.py introduced channel caching, but it had a side-effect of keeping channels open forever
In this PR we implement a process-wide ChannelCache object that keeps track of open channels. The code is largely inspired by https://github.com/grpc/grpc/blob/master/src/python/grpcio/grpc/_simple_stubs.py , but with the added support for grpc.aio channels.