Rao Shoaib wrote:
> Andrew Gallatin wrote:
>>
>> However, one of the primary uses of zero-copy sends is sendfile().
>> and thanks to 6459866, it is faster to copy. For context, see
>> http://markmail.org/message/oeowmlvfqi3kzttf
> Thanks for pointing this out -- we will look into this. However I
> suggest that you give sendfile another try, it has been updated to use
> vpm memory interfaces which allow larger mappings and are more efficient.
Using the same host, here is a variety of netperf -tTCP_SENDFILE
results using a 9000b MTU. The transmitter ran {Open,}Solaris on a Myri10GE
NIC in a very weak host (dual CPU 2.0 GHz "consumer grade" amd64).
The receiver was a very fast Linux host. I used a 9000b MTU
so as to make the lack of LSO in S10U1 matter less:
Solaris 10 1/06 s10x_u1wos_19a X86, GLD2 driver (no LSO available):
Recv Send Send Utilization Service
Demand
Socket Socket Message Elapsed Send Recv Send Recv
Size Size Size Time Throughput local remote local
remote
bytes bytes bytes secs. 10^6bits/s % M % S us/KB us/KB
87380 196608 196608 59.99 9900.34 32.71 11.00 0.541
0.364
OpenSolaris, 5.11 snv_114, GLDv2 driver (no LSO available):
87380 196608 196608 60.00 9900.01 64.08 11.15 1.060
0.369
OpenSolaris, 5.11 snv_114, GLDv3 driver (LSO disabled):
87380 196608 196608 60.00 9331.18 79.84 10.14 1.402
0.356
OpenSolaris, 5.11 snv_114, GLDv3 driver (LSO):
87380 196608 196608 60.46 9806.08 57.85 11.68 0.967
0.390
So it looks like there have been substantial performance regressions
since S10U1 which remain to be addressed. For comparision, here's
linux on the same host:
Linux, 2.6.22 (TSO disabled):
87380 65536 65536 30.02 9909.72 37.00 11.29 0.612
0.373
Linux, 2.6.22 (TSO in use):
87380 65536 65536 60.02 9900.70 27.61 11.31 0.457
0.374
Drew
_______________________________________________
networking-discuss mailing list
[email protected]
