guillaumeiooss
commented
3 years ago Hello,
In order to facilitate the investigation of this issue, I have isolated and simplified our benchmarking code for oneDNN from the rest of our tool and I have pushed it in the following repository: https://github.com/guillaumeiooss/onednn_meas_issue_July21
By default, this code corresponds to "Papi with flush", but you should be able to easily obtain the other measures (ex: by commenting the "flush_cache()" in "oneDNN_conv.c", or by switching the timing function used)
If you have any additional questions about how we measure our benchmarks, feel free to contact us ( nicoTolly or myself ). Thanks.
vpirogov
alheinecke
bgoglin
Summary
We tried to reproduce the results given in "Anatomy of High-Performance Deep Learning Convolutions on SIMD Architecture", Evangelos Georganas, Sasikanth Avancha, Kunal Banerjee, Dhiraj Kalamkar, Greg Henry, Hans Pabst and Alexander Heinecke. ArXiv:1808.05567v2 (20 August 2018) And did not get the expected results.
Version
Report oneDNN version and githash. Version information is printed to stdout in verbose mode. version 2.3 oneDNN commit a4c31c17ce4eb846bd1e80c0ff7a9699d37c3813
Environment
cmake 3.13.4-1 gcc 8.3.0-6 make 4.2.1-1.2
CMake log : see attachment
uname : Linux dahu-29.grenoble.grid5000.fr 4.19.0-16-amd64 #1 SMP Debian 4.19.181-1 (2021-03-19) x86_64 GNU/Linux
lscpu : Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Byte Order: Little Endian Address sizes: 46 bits physical, 48 bits virtual CPU(s): 64 On-line CPU(s) list: 0-63 Thread(s) per core: 2 Core(s) per socket: 16 Socket(s): 2 NUMA node(s): 2 Vendor ID: GenuineIntel CPU family: 6 Model: 85 Model name: Intel(R) Xeon(R) Gold 6130 CPU @ 2.10GHz Stepping: 4 CPU MHz: 2053.353 CPU max MHz: 3700.0000 CPU min MHz: 1000.0000 BogoMIPS: 4200.00 Virtualization: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 1024K L3 cache: 22528K NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62 NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single pti intel_ppin ssbd mba ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req pku ospke md_clear flush_l1d
Observed behavior
Hello,
We are currently working on benchmarking oneDNN on a set of convolutions and we ran into some problems. Basically, we struggle to reproduce the results given in [Georganas18]. Our architecture is Intel(R) Xeon(R) Gold 6130 (Skylake).
We tried two different methodologies :
We were surprised by the results for two reasons : 1) First the results are inconsistent between the two methods. Our benchmarking tool gets nearly always significantly better results than benchDNN.
2) Second, even with our method (the fastest one) we found the performance of oneDNN to be much lower than we would have expected, especially for a sequential implementation. Thus, we are currently at loss to explain the discrepancies in performance and we suspect that we are missing a crucial detail.
[Georganas18] "Anatomy of High-Performance Deep Learning Convolutions on SIMD Architecture", Evangelos Georganas, Sasikanth Avancha, Kunal Banerjee, Dhiraj Kalamkar, Greg Henry, Hans Pabst and Alexander Heinecke. ArXiv:1808.05567v2 (20 August 2018). In particular, we looked at Figure 4 in order to get an idea on where the performance of oneDNN should be.
Description of the evaluation methodology
We evaluate oneDNN on a set of convolution taken from classical deep neural network benchmarks, namely ResNet18, Yolo9000 and MobileNet. We only evaluate our benchmarks in inference mode, that is, with batches of size 1, and in the sequential case.
For benchDNN, we use the following command (build option: DNNL_CPU_RUNTIME=SEQ ) :
Example of problem description : ic3ih544iw544_oc32oh544ow544_kh3kw3_sh1sw1_n"yolo9000_0"
In our custom framework, we do not measure the time taken by initialization calls (to build the 'dnnl_primitive', plus the time taken by a 'write_to_dnnl_memory') but the time taken by the following portion of code :
CHECK(dnnl_primitive_execute( convolution_p, // Primitive to execute stream_cpu, // 3, // Number of elements in args conv_p_args // Arguments (data I/O) ));
CHECK(dnnl_primitive_execute( reorder_Output_p, // Primitive to execute stream_cpu, // 2, // Number of elements in args reorder_Output_p_args // Arguments (data I/O) ));
We ran both a code with this flush (cold cache) and without, and report both situations.
Measured performances
The measured results are reported in the chart (joined to this email), in percentage of the peak performance: -> The blue bar is the time reported by our framework by using "clock_gettime" and with cache flush (cold cache) -> The red bar is the time reported by our framework by using "clock_gettime" and without cache flush
-> The green bar is the time reported by our framework by using PAPI and with cache flush (cold cache) -> The violet bar is the time reported by our framework by using PAPI and without cache flush
-> The black bar is the time reported by benchDNN.
As you can see on the chart, the results are quite unexpected : -> The results of benchDNN is at most at 55% of the machine peak, which is much lower than the performances measured by our framework. -> In any case, all these measurements (both benchDNN and our framework) seems to be below their expected value (from [Georganas18]).
Therefore: -> Could you precise the evaluation methodology from which you obtained the performance measurement from [Georganas18] ? In particular, are you assuming hot or cold cache?
-> Do you have an idea on why benchDNN measured performance are that low in sequential?
-> In general, if you see a detail that we might have been missing, could you report it? Because, we are currently at loss to explain these discrepancy of performance.
We will be available, if you wish to have more details on our evaluation methodology, or if you wish for us to run additional experiments to refine our observations.
Thanks
CMakeOutput.log