penberg / posixbench

POSIX latency and energy-efficiency benchmark suite

POSIX Benchmarks

Getting Started

You need to configure your machine before running the benchmarks.

Disable TurboBoost.

sudo ./scripts/turbo-boost.sh

Disable CPU power saving mode:

sudo cpupower --cpu all frequency-set --governor performance

Configure huge pages:

sudo ./scripts/hugepages.sh

Kill background processes and services.

Benchmarking

Prerequisites:

• matplotlib
• pandas

make build
make bench

posixbench's Issues

Measure energy consumption

Let's see if we can use Intel's RAPL interface to measure energy consumption of POSIX operations. For example, we can query the PKG_ENERGY_STATUS MSR register before an operation and after an operation and calculate the amount of energy consumed for a single operation.

More information on the RAPL interface here:

http://web.eece.maine.edu/~vweaver/projects/rapl/

and an example program to access the MSRs here:

http://web.eece.maine.edu/~vweaver/projects/rapl/rapl-read.c

Machine topology discovery

We currently vary the placement of threads on each logical CPU core. Instead, we should use something like hwloc library to discover machine topology and run some of the benchmarks (for example eventfd) in the following configurations: socket local, NUMA local, and NUMA remote.

Ensure measuring thread state is placed on correct CPU

We currently allocate measuring thread state in the benchmark harness thread. Let's move the state to the CPU that's running the measuring thread.

POSIX IPC benchmarks

POSIX virtual memory benchmarks

Let's add benchmarks for mmap, mprotect, and others.

Avoid measurement bias

Let's analyze what measurement bias the posixbench suite has and address them.

The following paper is a good starting point for understand measurement bias:

Todd Mytkowicz, Amer Diwan, Matthias Hauswirth, and Peter F. Sweeney. 2009. Producing wrong data without doing anything obviously wrong! SIGARCH Comput. Archit. News 37, 1 (March 2009), 265–276.

Another data point is the causal profiling stuff from Emery Berger and others.

Large page benchmarks

The page fault benchmark uses 4K pages. Let's add a variant that uses large pages too.