Skip to contentSkip to conversation

raid5 xor functions and caches

DISCUSSION Andi Kleen 1 participant 1 message Mar 21, 2003

lore
Share & tools
lore

Andi Kleen

author
#0 lore raw

Hi,

I tuned the xor.h functions for x86-64 a bit. One of the changes I did was to use streaming stores, which are usually faster. But it's actually slower in the test it does a booting because that tests cache hot behaviour. NTI will invalidate the cache line, so it's eating a lot of cache misses.

Does it make sense to test cache hot behaviour? iirc RAID checksumming should be done without polluting caches. I'm considering to change it to benchmark a few MB of data buffers to avoid caching effects.

Also I'm not quite sure about the logic between XOR_SELECT_TEMPLATE prefering SSE :- if the CPU supports SSE2 then all functions including integer can be compiled with NTA prefetches and NTI stores (on some CPUs an integer NTI store is not a good idea though) For distribution kernels
they can be duplicated. If integer is really faster for some things it should be used this way.

Another thing I noticed is that the prefetch distance for the SSE functions is a bit too short. On modern CPUs 256 bytes is not long enough (on P4 that would be only two cachelines), it needs more (3-4 cachelines at least).

-Andi