Poor performance of the libmxnet if OMP_PLACES environment variable is present #14087
Comments
|
Hey, this is the MXNet Label Bot. |
|
@mxnet-label-bot add [bug] |
|
@mxnet-label-bot add [bug, performance, openmp] |
|
@mxnet-label-bot add [Environment Variables] |
|
@stsukrov could you try the below environment?
|
|
@pengzhao-intel Tried. It's issue about having an regression, IF something is present, not about existing good configuration. My assumption: it's caused by multiple OMPs, since it's gone as soon as I symlink gomp to iomp5. To my knowledge, (source: https://www.youtube.com/watch?v=Y7eVVfyo0U4, Michael Klemm) KMP_AFFINITY=granularity=fine,compact,1,0 is roughly OMP_PLACES=cores Right? Recall, I'm talking about 10x performance regression. |
|
@stsukrov thanks for the analysis. In step 3, two env works well in myside and could you double check? Hopes this can help to unblock your works during we are debugging the OMP_PLACES issue. ==============================================================
|
|
@pengzhao-intel NNVM_EXEC_MATCH_RANGE and MXNET_SUBGRAPH_BACKEND are additional optimizations. My results:
Note the message about ignoring OMP_PLACES.
Looks good to me. So, yes. just OMP_NUM_THREAD and KMP_AFFILITY do. OMP_PLACES kills the performance. |
|
Now, I'm doing a symlink And OMP_PLACES is not a trouble maker anymore. So, I interpret it as sign, that the issue is caused by 2 conflicting OMP runtimes. Quoting from: /~https://github.com/intel/mkl-dnn:
Yes, I know, that Intel OpenMP exports both GOMP_ and kmpc_ functions, but something seems to clash. |
|
@pengzhao-intel |
|
@stsukrov I am a little busy this week and I will look into the issue later. |
|
@pengzhao-intel |
|
@pengzhao-intel - Any suggestion here? Thank you. |
|
Sorry for the delay. @NeoZhangJianyu @yinghu5 will help to figure out the reason and fix the issue :) |
|
@sandeep-krishnamurthy For MXNet installed by PIP/Conda by now:
For MXNet installed by source:
Above setting after compile could avoid the issue of conflicting OMP runtimes. In you want to use OMP_PLACES to improve the performance in this case:
|
|
Thanks for providing work around recommendations. Was actually dreaming of a build, that just works without venv or LD_LIBRARY_PATH-thaumaturgy. |
|
@stsukrov In development server, we always install many many different release libs. So the conflict will happen easily. venv and LD_LIBRARY_PATH are easy way to avoid/clear the conflict of lib files. In server for deployment, it's better to copy the libs in /usr/lib if the install/upgrade tool/script are good enough. venv: we could remove totally if we can't resolve the conflict. LD_LIBRARY_PATH:
|
|
Just to make a few connections between tickets: Here was an attempt to remove bundled conflicting OpenMP library: #12160 Link to the stale dev list discussion. A proposal from my side for a robust and stable solution is to make a openmp environment checking script that would:
|
Description
Default build depends on both GOMP and Intel OMP runtime.
Environment info (Required)
----------Python Info----------
Version : 3.5.2
Compiler : GCC 5.4.0 20160609
Build : ('default', 'Nov 12 2018 13:43:14')
Arch : ('64bit', 'ELF')
------------Pip Info-----------
Version : 8.1.1
Directory : /usr/lib/python3/dist-packages/pip
----------MXNet Info-----------
Version : 1.5.0
Directory : /home/stsukrov/workspace/mxnet/python/mxnet
Hashtag not found. Not installed from pre-built package.
----------System Info----------
Platform : Linux-4.4.0-1074-aws-x86_64-with-Ubuntu-16.04-xenial
system : Linux
node : ip-172-31-1-212
release : 4.4.0-1074-aws
version : #84-Ubuntu SMP Thu Dec 6 08:57:58 UTC 2018
----------Hardware Info----------
machine : x86_64
processor : x86_64
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 72
On-line CPU(s) list: 0-71
Thread(s) per core: 2
Core(s) per socket: 18
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Platinum 8124M CPU @ 3.00GHz
Stepping: 4
CPU MHz: 3000.000
BogoMIPS: 6000.00
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 32K
L1i cache: 32K
L2 cache: 1024K
L3 cache: 25344K
NUMA node0 CPU(s): 0-17,36-53
NUMA node1 CPU(s): 18-35,54-71
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc aperfmperf pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single kaiser fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx avx512f rdseed adx smap clflushopt clwb avx512cd xsaveopt xsavec xgetbv1 ida arat pku
----------Network Test----------
Setting timeout: 10
Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0155 sec, LOAD: 0.0708 sec.
Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0348 sec, LOAD: 0.2342 sec.
Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0010 sec, LOAD: 0.3944 sec.
Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.0896 sec, LOAD: 0.7671 sec.
Timing for MXNet: /~https://github.com/apache/incubator-mxnet, DNS: 0.0006 sec, LOAD: 0.9998 sec.
Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.2689 sec, LOAD: 1.0834 sec.
Compiler (gcc/clang/mingw/visual studio):
gcc-5
MXNet commit hash: d684c59
Build info (Required if built from source)
Build config:
make -j20
Minimum reproducible example
Get mxnet
Build from scratch
run the benchmark - get performance on mxnet/batch size=1 like 200img/sec
set OMP_PLACES=cores
rerun the benchmark
get performance like 40 - which is poor
Steps to reproduce
(Paste the commands you ran that produced the error.)
Rerunning with
export OMP_DISPLAY_ENV=verbose we have:
OPENMP DISPLAY ENVIRONMENT BEGIN
_OPENMP = '201511'
OMP_DYNAMIC = 'FALSE'
OMP_NESTED = 'FALSE'
OMP_NUM_THREADS = '72'
OMP_SCHEDULE = 'DYNAMIC'
OMP_PROC_BIND = 'TRUE'
OMP_PLACES = '{0,36},{1,37},{2,38},{3,39},{4,40},{5,41},{6,42},{7,43},{8,44},{9,45},{10,46},{11,47},{12,48},{13,49},{14,50},{15,51},{16,52},{17,53},{18,54},{19,55},{20,56},{21,57},{22,58},{23,59},{24,60},{25,61},{26,62},{27,63},{28,64},{29,65},{30,66},{31,67},{32,68},{33,69},{34,70},{35,71}'
OMP_STACKSIZE = '0'
OMP_WAIT_POLICY = 'PASSIVE'
OMP_THREAD_LIMIT = '4294967295'
OMP_MAX_ACTIVE_LEVELS = '2147483647'
OMP_CANCELLATION = 'FALSE'
OMP_DEFAULT_DEVICE = '0'
OMP_MAX_TASK_PRIORITY = '0'
GOMP_CPU_AFFINITY = ''
GOMP_STACKSIZE = '0'
GOMP_SPINCOUNT = '300000'
OPENMP DISPLAY ENVIRONMENT END
OPENMP DISPLAY ENVIRONMENT BEGIN
_OPENMP='201611'
[host] KMP_ABORT_DELAY='0'
[host] KMP_ADAPTIVE_LOCK_PROPS='1,1024'
[host] KMP_ALIGN_ALLOC='64'
[host] KMP_ALL_THREADPRIVATE='288'
[host] KMP_ATOMIC_MODE='2'
[host] KMP_BLOCKTIME='200'
[host] KMP_CPUINFO_FILE: value is not defined
[host] KMP_DETERMINISTIC_REDUCTION='FALSE'
[host] KMP_DEVICE_THREAD_LIMIT='2147483647'
[host] KMP_DISP_HAND_THREAD='FALSE'
[host] KMP_DISP_NUM_BUFFERS='7'
[host] KMP_DUPLICATE_LIB_OK='FALSE'
[host] KMP_FORCE_REDUCTION: value is not defined
[host] KMP_FOREIGN_THREADS_THREADPRIVATE='TRUE'
[host] KMP_FORKJOIN_BARRIER='2,2'
[host] KMP_FORKJOIN_BARRIER_PATTERN='hyper,hyper'
[host] KMP_FORKJOIN_FRAMES='TRUE'
[host] KMP_FORKJOIN_FRAMES_MODE='3'
[host] KMP_GTID_MODE='3'
[host] KMP_HANDLE_SIGNALS='FALSE'
[host] KMP_HOT_TEAMS_MAX_LEVEL='1'
[host] KMP_HOT_TEAMS_MODE='0'
[host] KMP_INIT_AT_FORK='TRUE'
[host] KMP_INIT_WAIT='2048'
[host] KMP_ITT_PREPARE_DELAY='0'
[host] KMP_LIBRARY='throughput'
[host] KMP_LOCK_KIND='queuing'
[host] KMP_MALLOC_POOL_INCR='1M'
[host] KMP_NEXT_WAIT='1024'
[host] KMP_NUM_LOCKS_IN_BLOCK='1'
[host] KMP_PLAIN_BARRIER='2,2'
[host] KMP_PLAIN_BARRIER_PATTERN='hyper,hyper'
[host] KMP_REDUCTION_BARRIER='1,1'
[host] KMP_REDUCTION_BARRIER_PATTERN='hyper,hyper'
[host] KMP_SCHEDULE='static,balanced;guided,iterative'
[host] KMP_SETTINGS='FALSE'
[host] KMP_SPIN_BACKOFF_PARAMS='4096,100'
[host] KMP_STACKOFFSET='64'
[host] KMP_STACKPAD='0'
[host] KMP_STACKSIZE='4M'
[host] KMP_STORAGE_MAP='FALSE'
[host] KMP_TASKING='2'
[host] KMP_TASKLOOP_MIN_TASKS='0'
[host] KMP_TASK_STEALING_CONSTRAINT='1'
[host] KMP_TEAMS_THREAD_LIMIT='72'
[host] KMP_TOPOLOGY_METHOD='all'
[host] KMP_USER_LEVEL_MWAIT='FALSE'
[host] KMP_VERSION='FALSE'
[host] KMP_WARNINGS='TRUE'
[host] OMP_AFFINITY_FORMAT='OMP: pid %P tid %T thread %n bound to OS proc set {%a}'
[host] OMP_ALLOCATOR='omp_default_mem_alloc'
[host] OMP_CANCELLATION='FALSE'
[host] OMP_DEFAULT_DEVICE='0'
[host] OMP_DISPLAY_AFFINITY='FALSE'
[host] OMP_DISPLAY_ENV='VERBOSE'
[host] OMP_DYNAMIC='FALSE'
[host] OMP_MAX_ACTIVE_LEVELS='2147483647'
[host] OMP_MAX_TASK_PRIORITY='0'
[host] OMP_NESTED='FALSE'
[host] OMP_NUM_THREADS: value is not defined
[host] OMP_PLACES='cores'
[host] OMP_PROC_BIND='spread'
[host] OMP_SCHEDULE='static'
[host] OMP_STACKSIZE='4M'
OMP_TARGET_OFFLOAD=DEFAULT
[host] OMP_THREAD_LIMIT='2147483647'
[host] OMP_TOOL='enabled'
[host] OMP_TOOL_LIBRARIES: value is not defined
[host] OMP_WAIT_POLICY='PASSIVE'
[host] KMP_AFFINITY='noverbose,warnings,respect,granularity=core,compact,0,0'
OPENMP DISPLAY ENVIRONMENT END
The first one is probably GOMP
The second one is IOMP
Setting KMP_AFFINITY=verbose gives this:
OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.
OMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info
OMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0,36
OMP: Info #156: KMP_AFFINITY: 2 available OS procs
OMP: Info #157: KMP_AFFINITY: Uniform topology
OMP: Info #179: KMP_AFFINITY: 1 packages x 1 cores/pkg x 2 threads/core (1 total cores)
OMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:
OMP: Info #171: KMP_AFFINITY: OS proc 0 maps to package 0 thread 0
OMP: Info #171: KMP_AFFINITY: OS proc 36 maps to package 0 thread 1
which may mean Intel OMP suddenly thinks, it's just one core
Without OMP_PLACES looks fine:
OMP: Info #157: KMP_AFFINITY: Uniform topology
OMP: Info #159: KMP_AFFINITY: 2 packages x 18 cores/pkg x 2 threads/core (36 total cores)
What have you tried to solve it?
The text was updated successfully, but these errors were encountered: