From: Paul Mackerras <[EMAIL PROTECTED]> Date: Sat, 21 Jun 2008 14:30:02 +1000
> Is this application really transferring bulk data and using buffers > that aren't a multiple of the page size? Do you know whether the > copies ended up being misaligned? We used to cache align the sub-buffers carved out of the per-socket anonymous buffer page that TCP, UDP, and other protocols use to collect user write()/sendmsg() data when the outgoing interface supports scatter-gather and checksumming (basically just about any reasonable ethernet device these days). But that alignment was removed from net/ipv4/tcp.c (I forget the exact reasons, I think it was space wastage). The net result is that you will therefore see a lot of misaligned copies for networking sends these days. In my opinion optimizing mempcy/user-copy in the most general way is the best course of action. Whatever histogram of sizes and alignments you obtain with sampling today will change unpredictably in the future. I'm saying this as someone who fell into that trap on sparc64 several times in the past, and I have 5 memcpy/user-copy/memset variants to maintain on that platform these days :-/ _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@ozlabs.org https://ozlabs.org/mailman/listinfo/linuxppc-dev
