Revision a6e13ce96934354f9bae2fbe134071852e6340ee authored by Junjie Qian on 13 December 2016, 21:58:26 UTC, committed by Junjie Qian on 25 February 2017, 18:53:48 UTC
Packing aggregated matrixes into continuous buffer so less MPI operations are used.

Usage:
Introduce a configurable threshold size (units = KB) and only pack
gradients into a continous buffer whose total size is below the threshold.
This is to avoid possible memory overflow.

The default threshold size is 32KB. To change this value, specify
"packThresholdsizeInKB=value" (e.g. 2048 for 2MB) in train/eval config
file.

Notes:
acquire continous buffer only if NCCL is not supported

Fallback to normal execution if not enough extra continous memory available

change reduce all to unblocking reduce all

SimpleDistAggregate: Broadcast the aggregated headers to all nodes

Configurable threshold size to pack gradients in a buffer

simpleAggregator: pack the gradient into continous buffer if its size is less than PackThresholdSize

Add all constant definitions to a header file in Common
1 parent 9375555
History
File Mode Size
Dependencies
Documentation
Examples
Scripts
Source
Tests
Tools
Tutorials
bindings
.clang-format -rw-r--r-- 931 bytes
.gitattributes -rw-r--r-- 2.5 KB
.gitignore -rw-r--r-- 5.3 KB
.gitmodules -rw-r--r-- 211 bytes
CNTK.Common.props -rw-r--r-- 1.2 KB
CNTK.Cpp.props -rw-r--r-- 7.7 KB
CNTK.sln -rw-r--r-- 182.5 KB
CONTRIBUTING.md -rw-r--r-- 199 bytes
CppCntk.vssettings -rw-r--r-- 10.0 KB
LICENSE.md -rw-r--r-- 4.6 KB
Makefile -rw-r--r-- 51.3 KB
README.md -rw-r--r-- 9.5 KB
configure -rwxr-xr-x 31.9 KB

README.md

back to top