| Commit message (Collapse) | Author | Age |
|
|
|
|
|
|
|
|
|
|
|
|
| |
more than one device-to-device copy stream per GPU device.
This is an experimental feature that will have no effect unless
copy operations explicitly request a stream other than 0, which
currently does not occur anywhere in a standard build.
Eventually it may be of benefit in the presence of multiple
bi-directional concurrent data copies.
PiperOrigin-RevId: 202354513
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Without this change, a shared resource (e.g. an Iterator) could not be
created in one session `s1`, and used in a later session `s2` after
`s1` was closed, because the iterator might indirectly capture devices
from the previous session, and use them after they are freed when the
`WorkerSession` was deleted.
The current change only affects the singleton "legacy" WorkerSession,
which is never deleted, but this is necessary to switch all sessions
to use separate WorkerSession objects.
PiperOrigin-RevId: 193541426
|
|
|
|
|
|
|
|
| |
tensorflow::str_util equivalents.
This will allow the deprecated methods to be removed.
PiperOrigin-RevId: 191350894
|
|
|
|
|
|
|
|
|
|
| |
Adds a test in the variant_op_copy_test.
Modifies the base GPUDevice to use this registry if it sees a singleton variant.
Modifies the rendezvous manager to do the same.
PiperOrigin-RevId: 170908757
|
|
|
|
| |
PiperOrigin-RevId: 167401527
|
|
|
|
| |
PiperOrigin-RevId: 162782660
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The base_rendezvous_mgr handles transferring a tensor using DMAs to non-"host"
devices such as GPUs in the SameWorkerRecvDone function. This function would
use the worker_env's device_mgr to obtain a pointer to the relevant Device
(using the LookupDevice call). In the ClusterSpec-propagation world, using the
environment's device_set means that the devices are not renamed, often
resulting in devices not being found.
This change updates BaseRemoteRendezvous to use the WorkerSession stored when
the BaseRemoteRendezvous is initialized. The WorkerSession has a pointer to a
DeviceMgr that contains the appropriately renamed devices for the given
session the Rendezvous is associated with.
Note: because we have a fast-path host-device-only copy, the original bug does
not show up when using 2 CPU devices. I have added a test to ensure that
transferring between 2 CPU devices works in a ClusterSpec propagation session,
but note that this test does not actually reproduce the motivating bug.
In the process of writing a test for the original bug, I discovered another
latent bug in ClusterSpec propagation where if there were 2 CPU devices
(i.e. due to explicit server configuration to have 2 CPU devices), a DCHECK
could be triggered. The Master::CreateSession would call
`device_set->set_client_device` multiple times (once for each CPU device).
PiperOrigin-RevId: 162680163
|
|
|
|
|
|
|
|
|
|
|
| |
ClusterSpec propagation is a capability upgrade for TensorFlow that should make
it much easier to (1) build distributed TensorFlow clusters, and (2) handle
node failures. The ClusterSpec propagation capability allows TensorFlow workers
to be booted independently of each other, and with no knowledge about others.
The client can then construct a ClusterDef (ClusterSpec), and then send it
to the TF master at session creation. The master in turn then propagates the
ClusterDef along to all of the workers.
Change: 155159972
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
State in workers is currently splayed across graph_mgr, rendezvous_mgr, and
additional components. This has resulted in it being difficult to ensure proper
cleanup and shut down of the worker components.
In addition to paving the way for a more reliable shut down, this CL also sets
up the beginnings of ClusterSpec propagation.
ClusterSpec propagation is a capability upgrade for TensorFlow that should make
it much easier to (1) build distributed TensorFlow clusters, and (2) handle
node failures. After the ClusterSpec propagation capability is fully
implemented, the TensorFlow workers can be booted independently of each other,
and with no knowledge about others. A client can then query a central cluster
scheduler or other API to find all of the workers, and then send the
ClusterDef (ClusterSpec) to the TF master, which then propagates that along to
all of the workers.
This change is only the first of a sequence to fully implement ClusterSpec
propagation in TensorFlow.
Change: 151229111
|
|
|
|
| |
Change: 135518672
|
|
|
|
|
|
|
| |
For example:
- Unused variables/functions/typedefs
- Thread-safety annotations
Change: 133986225
|
|
|
|
|
|
|
| |
-Use std::move when assigning std::function to reduce some simple allocations.'
-Use std::bind to avoid copy of std::function in lambda statements.
Change: 129654152
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
done by models distributed across many devices. A small
microbenchmark model that runs two banks (A and B) of 30 nodes with a
30x30 full shuffle between them, where each of the nodes in A and in B
run with one node on each of the 30 devices (so 30*29+30+30, or ~930
separate RPCs) was showing ~111,000 allocations per iteration of the graph.
With the changes here, this is now down to ~64,300 allocations per iteration.
Changes include:
o DeviceContext::CopyDeviceTensorToCPU and related helper routines:
use StringPiece instead of const string& for the tensor name (avoids
creating a string in some cases where the caller only has a
StringPiece available).
o Change some Rendezvous and BaseRemoteRendezvous interfaces to
take a 'const Rendezvous::ParsedKey& key', rather than 'const string& key'.
In many cases, the callers were already having to parse the key
into a ParsedKey, and so we were doing the parsing multiple times at
different levels as we processed receiving or sending of a tensor. This
reduces the number of times that we parse a key as it flows from a Send
node through to a Recv node on another worker.
o Changed Rendezvous::ParsedKey so that it makes a copy of the underlying
full key, and then uses StringPiece objects to point into this copy for
the src_device, dst_device, and edge_name pieces. This turns 3 string
allocations into 1 per Rendezvous::ParseKey call.
o Added new StringPiece Rendezvous::ParsedKey::FullKey() accessor to
return a StringPiece for the underlying full key, and used that in a
few places (mostly logging) where that is useful.
o In many places, used std::move(function_variable) when assigning to
an instance variable. This eliminates a very large number of excess
std::function allocations/initializations (~56000 of the baseline
allocations were related to std::function setup or cloning, and this
is now down to ~11000 after this cl).
o In the RPC-based remote workers (StubbyRemoteWorker and
GrpcRemoteWorker), changed the code path in RecvTensorAsync to avoid
creation of a std::function with 6 arguments unless necessary. There
are three cases now handled separately:
(a) We're not logging, and we didn't make a copy of the request that we
need to free: just use the passed in 'StatusCallback done' object
directly, without creating a wrapper std::function object at all
(b) We're not logging, but we made a copy of the request that we
need to free: we create a simple wrapper std::function that
invokes the passed in 'done' callback, and then frees the
req_copy request copy object.
(c) We're logging: we create the std::function object with all the
necessary state to log when the recv has finished.
o Changed DeviceMgr::LookupDevice to take a StringPiece, rather than a
const string&, and changed the hash table to use StringPiece keys.
This allows clients that just have a StringPiece device name in their
hand to avoid a string creation to lookup the Device* object.
o Changed ExecutorState to use a specialized TaggedNodeReadyQueue that
internally uses a gtl::InlinedVector<TaggedNode, 16>, rather than
using a std::deque<TaggedNode> for keeping track of nodes ready to
execute. This is faster because it avoids allocations entirely if the
ready node queue doesn't get bigger than 16, and inlined vectors are
generally faster than std::deque, at a minor risk of using more memory
if this queue grows to very large numbers of ready nodes (mostly imaginable
only in pathological graphs).
o In ExecutorState::Process, allocated a single ExecutorState::AsyncState
object to keep track of all the state we need to preserve for an asynchronously
executed node, rather than keeping this state implicitly via a very large
number of arguments to a lamda function.
o Added new atomic std::atomic<bool> status_is_ok_ in
BaseRemoteRendezvous. This allows us to avoid acquiring the lock when
we just want to check if the status is non-OK in
BaseRemoteRendezvous::Send and BaseRemoteRendezvous::ValidateDevices.
o In GraphMgr::RunAllDone, changed assignment of args.runner to avoid
one extra level of std::function indirection (binding the function directly
to the ThreadPool::Schedule routine, rather than creating an intermediate
lambda function that invokes this inside the body of the lambda.
o Added freelist of RpcRecvTensorCall objects in
third_party/tensorflow/core/distributed_runtime/rpc/rpc_rendezvous_mgr.cc
o Changed third_party/tensorflow/core/framework/rendezvous.cc to keep the
hashtable of Item* objects keyed by uint64 (hash of the tensor name), rather
than the full-string tensor name. Collisions in the 64-bit hash space
should basically never happen.
o Sped up DeviceNameUtils::ParseFullName by optimizing for the common
ordering of parts of /job, /replica, /task, /device. The parsing code
was general enough to handle any order, but did so by comparing the
prefixes 4, 3, 2, and 1 times, respectively, rather than 1, 1, 1, and 1 times.
o Sped up DeviceNameUtils::SplitDeviceName to avoid extra string copies.
Change: 125991891
|
|
|
|
| |
Change: 123900938
|
|
|
|
|
|
| |
distributed_runtime to common_runtime in prepration for being shared by
DirectSession.
Change: 121339508
|
|
This includes a gRPC server (grpc_tensorflow_server) that can serve as both
the master of a distributed TensorFlow computation, and an individual worker
in the computation. The GrpcSession class is included to allow client programs
(including Python clients) to interact with a server.
See tensorflow/core/distributed_runtime/README.md for usage instructions.
This change partially addresses issue #23.
Change: 115634191
|