[SYSTEMDS-2487] Native Dnn operations crashing in over-provisioned parfor

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: SystemML 1.2
• Assignee: Matthias Boehm
• Reporter: Matthias Boehm
• Created: 08/Aug/18 21:19
• Updated: 08/Aug/18 21:37
• Resolved: 08/Aug/18 21:37

Description

In case parfor does not consume all the available parallelism, we propagate this parallelism down to individual operations with slight (max 50%) overprovisioning. For example, if we have 80vcores, and parfor is assigned k=47, we still assign k=2 to individual operations.

However, with native DNN operations this causes JVM crashes as follows:

#
# A fatal error has been detected by the Java Runtime Environment:
#
#  SIGFPE (0x8) at pc=0x00007f5de21902d6, pid=335027, tid=0x00007f5df8bcb700
#
# JRE version: OpenJDK Runtime Environment (8.0_161-b14) (build 1.8.0_161-b14)
# Java VM: OpenJDK 64-Bit Server VM (25.161-b14 mixed mode linux-amd64 )
# Problematic frame:
# C  [libmkl_avx512.so+0x206d2d6][thread 140041622857472 also had an error]
  mkl_dnn_avx512_bkdGemmDirectConv_F64+0x276

Hence, when native BLAS or DNN libraries are loaded, we should be more conservative and not over-provision at all.