From: Daniel Borkmann > On 02/24/2014 11:17 AM, David Laight wrote: > > From: Daniel Borkmann > >> Drivers are allowed to set NETIF_F_SCTP_CSUM if they have > >> hardware crc32c checksumming support for the SCTP protocol. > >> Currently, NETIF_F_SCTP_CSUM flag is available in igb, > >> ixgbe, i40e/i40evf drivers and for vlan devices. > >> > >> If we don't have NETIF_F_SCTP_CSUM then crc32c is done > >> through CPU instructions, invoked from crypto layer, or > >> if not available as slow-path fallback in software. > >> > >> Currently, loopback device propagates checksum offloading > >> feature flags in dev->features, but is missing SCTP checksum > >> offloading. Therefore, account for NETIF_F_SCTP_CSUM as > >> well. > >> > >> Before patch: > >> > >> ./netperf_sctp -H 192.168.0.100 -t SCTP_STREAM_MANY > >> SCTP 1-TO-MANY STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to 192.168.0.100 () port 0 AF_INET > >> Recv Send Send > >> Socket Socket Message Elapsed > >> Size Size Size Time Throughput > >> bytes bytes bytes secs. 10^6bits/sec > >> > >> 4194304 4194304 4096 10.00 4683.50 > >> > >> After patch: > > ... > >> 4194304 4194304 4096 10.00 15348.26 > > > > That seems a much larger increase than you'd expect from removing > > a software CRC of the data chunks. > > Are you sure that some other difference in the data flows wasn't > > also triggered. > > Yes, I run this multiple times with similar results and I double-checked > it with perf. Current code triggers crc32c implementation in software > fallback on my machine which is very expensive. I'm sure it shouldn't be that expensive, you are implying that it spent about 70% of the time doing crc32. The loop should be dominated by the per-byte lookup in a 256 word table. With 4k data the table will soon be in the data cache. Unless it is (stupidly) generating the table on each call, or trying to use a crc32 instruction, faulting, and emulating it, I wouldn't really have expected more than a few % improvement. > > I'm also not sure that 4096 is a representative message size for SCTP. > > I used netperf default in this case. I was just pointing out that the defaults may not be appropriate here. David -- To unsubscribe from this list: send the line "unsubscribe linux-sctp" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
