Phoronix article – clear linux benchmarks (2018-05-16)
Phoronix posted an article about Clear linux. The goal was to look for changes in the system after all packages were rebuilt with GCC 8.1.
While I don’t have the particular combination to reproduce, this posting goes through the workloads to better understand what is being measured.
Results are detailed in the table below. One of the more surprising results is that scimark2 was mostly flat in a previous phoronix gcc 8.1 article, yet shows improvements here. Otherwise it shows clearlinux as mostly faster after the change to 8.1 from 7.3.
| Benchmark | Phoronix observations | My observations | Analysis |
|---|---|---|---|
| ffte | FFTE performance improved slightly almost all systems. | Single threaded, runs for 3.5 seconds, some backend stalls. Mostly toy benchmark. | analysis |
| scimark2 | Scimark2 improved on all systems. | scimark2 performance was unchanged in previous benchmark run from gcc 7.3 to gcc 8.1, so unclear why it should improve here. | analysis |
| himeno | himeno improved on most systems. | Single threaded benchmark most limited by backend stalls. Himeno also showed performance improvements when recompiled from gcc 7.3 to 8.1. | analysis |
| 7-zip | Overall just slightly lower performance with ubuntu 18.04 | Multithreaded with high rate of speculative misses and some frontend stalls. | analysis |
| build-apache | Slower compile times when switched to gcc 8.1 | build-apache fails to build on ubuntu 18.04 due to unresolved references to SSLv3_client_method, SSLv2_client_method, CRYPTO_malloc_init. Google search suggests this problem was introduced in the 1.1.0 version of libopenssl. Most likely clearlinux has a fixed version. Pass on this until Phoronix fixes upstream benchmark. | |
| rbenchmark | Consistently slower in clearlinux with gcc 8. | Single threaded with some backend stalls; sub-phases to the workloads. | analysis |
| caffe | Builds slower when based on gcc 8.1 | Build dependencies in building caffe on libopencv-dev. Installed, bug google search suggests possibly version skew issues. | |
| tensorflow | No significant changes. | IPC around 1 with backend stalls. Some I/O. | analysis |
| redis: lpush | Small improvements on most systems. | Not CPU-bound (<1%) more latency. Runs in few seconds. | analysis |
| redis: set | Small improvements on most systems | Not CPU-bound (<1%) more latency. Runs in few seconds. | analysis |
| nginx | Mixed, some better and some worse. | Single-threaded server and single-threaded driver. More system time than user time. Some backend stalls. | analysis |
| apache | Small improvements on half the systems. | Multiple server threads and single-threaded driver. More system time than user time. Some backend stalls. Many frontend stalls. | analysis |
| phpbench | No significant changes. | Single-threaded with high IPC. Lower IPC on AMD than Intel. | analysis |
| scikit-learn | Small improvements on all systems. | Single-threaded; high IPC with some backend stalls. Python program. | analysis |
| n-queens | Small improvements on all systems. | On_Core 100%, with large number of branch misses. | analysis |
Comments
Phoronix article – clear linux benchmarks (2018-05-16) — No Comments
HTML tags allowed in your comment: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>