line - cache line size
[ -M <len>
] [ -W <warmups>
] [ -N
tries to determine the cache line size in bytes of the largest cache
which is smaller than len
creates pointer chains which access the first word on each cache
line on a page (randomly meandering through all the lines in a page before
jumping to the next page). It measures the average memory latency for a
variety of line sizes, starting with a line size of one word. When it finds an
increase in the average latency that is significantly larger than the latency
for the smaller line size then it assumes that it has found the line size.
This algorithm works because for line sizes less than the true line size, at
least two line
cache lines fit in the space of a true cache line. Since
that cache line will be accessed twice, the first access will cause an
expensive cache miss, while the second access will be a cache hit. Once the
cache line is equal to the true cache line size, then all accesses
will cause cache misses.
is an experimental benchmark, but it seems to work well on most
lmbench(8), tlb(8), cache(8), par_mem(8).
Carl Staelin and Larry McVoy
Comments, suggestions, and bug reports are always welcome.