Memory Bandwidth
Running the Script
docker run --rm --privileged ghcr.io/sparecores/benchmark:main bw_mem.sh
In production, the median run takes ~4.0 minutes. See
Run Duration for how it was measured, and
Production Setup for how production runs differ
from a plain docker run.
Workload
bw_mem.sh runs LMbenchLMbenchA suite of portable micro-benchmarks for Unix-like systems, written by Larry McVoy and Carl Staelin. Its tools each measure one basic operation, such as memory bandwidth and latency, context switches, or system call overhead.LMbenchGlossary's bw_membw_memThe memory bandwidth tool of LMbench. It moves data through a buffer of a given size in one of several modes, such as rd (read and sum an array), wr (write a constant to each element), rdwr (read, then write the same location), and the copy modes cp, fcp, and bcopy. It reports the buffer size in MB and the bandwidth in MB/s; with -P, it runs several processes, each with its own buffer, and reports their total bandwidth.`bw_mem` manualGlossary, from Ubuntu's lmbench
package, in three of its modes:
rdreads data into the processor, computing the sum of an array.wrwrites data to memory, assigning a constant value to each element.rdwrreads and then writes the same location, adding the current value to a running sum before assigning a new constant value.
Each mode accesses every fourth word of an array of integers. The script runs
each mode over the following 11 sizes, 33 runs in sequence: 16k, 256k,
1M, 2M, 4M, 8M, 16M, 32M, 64M, 256M, and 512M.
Each run is nice -n -20 bw_mem -P $(nproc) <size> <op>niceA Unix command that runs a program with a changed scheduling priority, from -20 (highest) to 19 (lowest). Raising a process's priority above the default needs the CAP_SYS_NICE capability, which root has.man pageGlossary. With -P, each of
the nproc processes allocates its own buffer of the given size, and the
reported bandwidth is the total across all processes.
The bw_mem manual
opens by describing a buffer twice the given size that is copied from one half
to the other. That description applies to the copy modes, which the image does
not run; rd, wr, and rdwr allocate one buffer of the given size.
bw_mem Version
The bw_membw_memThe memory bandwidth tool of LMbench. It moves data through a buffer of a given size in one of several modes, such as rd (read and sum an array), wr (write a constant to each element), rdwr (read, then write the same location), and the copy modes cp, fcp, and bcopy. It reports the buffer size in MB and the bandwidth in MB/s; with -P, it runs several processes, each with its own buffer, and reports their total bandwidth.`bw_mem` manualGlossary program comes from Ubuntu 24.04's lmbench package, whose version
may differ from the upstream LMbenchLMbenchA suite of portable micro-benchmarks for Unix-like systems, written by Larry McVoy and Carl Staelin. Its tools each measure one basic operation, such as memory bandwidth and latency, context switches, or system call overhead.LMbenchGlossary source.
Output
The bw_mem.sh script prints one line per successful run,
<op> <size> <bandwidth>, such as rd 0.016384 147885.61:
<op>isrd,wr, orrdwr.<size>is the per-process buffer size in MB of 10^6 bytes, so16kprints as0.016384.<bandwidth>is the total MB/s across all processes.
A size that fails or prints unexpected output is skipped without an error, so a run can have fewer than 33 lines.