Building & Installing Memory Benchmark

This uses autotools. If you're compiling from git, run autogen.sh and then make. Otherwise, use ./configure && make.

To install, run sudo make install.

To run the pmmeter program for measuring memory latency and throughput.

This software is derived from NumaTop source code. Please refer https://github.com/intel/numatop

pmmeter is a microbenchmark to measure memory latency and bandwidth which is derived from mgen test program of NumaTop source code.

common.h is derived from OptaneStudy project. Please refer https://github.com/NVSL/OptaneStudy

Build Dependencies

It requires following libraries:

• numactl-devel or libnuma-dev(el)
• libncurses
• libpthread
• gcc (version 7 or later)

Supported Kernels

The recommended kernel version is the latest stable kernel, currently 4.15.

The minimum kernel version supported is 3.16

For Haswell supporting, please also apply a perf patch on 3.16. The patch is kernel_patches/0001-perf-x86-Widen-Haswell-OFFCORE-mask.patch.

The patch can also be found at following link: http://www.gossamer-threads.com/lists/linux/kernel/1964864

Directories

src: pmmeter source code. pmmeter is a microbenchmark which can generate memory access with runtime latency value among CPUs.

common: common code for all platforms.

intel : Intel platform-specific code.

test : test codes

kernel_patches: the required kernel patches.

Supported Hardware

numatop is supported on Intel Xeon processors: 5500-series, 6500/7500-series, 5600 series, E7-x8xx-series, and E5-16xx/24xx/26xx/46xx-series.

E5-16xx/24xx/26xx/46xx-series had better be updated to latest CPU microcode (microcode must be 0x618+ or 0x70c+).

To learn about NumaTOP, please visit http://01.org/numatop

Examples

• print usage;
  ./pmmeter -h
• mount pmem device as an App Direct mode on numa 0 and numa 1 node resoectively.
  $ sudo mount -o dax /dev/pmem0p1 /mnt/pmem_fsdax0/ # on node 0
  $ sudo mount -o dax /dev/pmem1p1 /mnt/pmem_fsdax1/ # on node 1
• ./pmmeter

log format

• pattern name; [[bw]_[r|s]_][load|store|ntload|nstore]_[clflush|clwb|clflushopt]_[fence|nofence]_[movq|movd|movb]_[64|128|256]
  • bw: bandwidth
  • r: random access, s: sequential access
  • load: load, store: store, ntload: non temporal load, nstore: non temporal store
  • clflush: clflush, clwb: clwb, clflushopt: clflushopt
  • fence: with fence, nofence: without fence
  • movq: mov quad word, movd: mov dobule word, movb: mov byte
  • 64: 64 byte, 128: 128 byte, 256: 256 byte