This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] Improve string benchtest timing
On 22/05/19 4:41 PM, Wilco Dijkstra wrote:
> Well if I force the copies to be mutually unaligned, there is only about 1% difference
> for a few of the memcpy implementations compared to them being always aligned
> The others show identical performance whether aligned or not. This is not too
> surprising since the test is basically waiting for DRAM most of the time.
That's a good point. Is that the case for thunderx as well? IIRC they
perform particularly badly with misaligned code but I don't know if
they're bad enough to be significant in the face of DRAM waits.
> So if we wanted to measure something useful we'd need to do it differently. Maybe
> the goal was to measure DRAM bandwidth? If so we could modify it to compare
> copy bandwidth for just a few different sizes (corresponding with typical L1/L2/L3 sizes).
That sounds like a good idea.
Siddhesh