Re: Network latency regressions from 2.6.22 to 2.6.29 (results with IRQ affinity)

[Eric Dumazet <dada1@cosmosbay.com>]

Christoph Lameter a écrit :
> On Mon, 20 Apr 2009, Eric Dumazet wrote:
> 
>> No changes in udpping but noise.
> 
> No changes in the results of udpping that you would not consider noise?

Only changes where when every cpus played, so it really is not meaningfull.

> 
> The time syscalls are virtual syscalls which are quite cheap.

Not on 32bits afaik, with glibc I use at least, strace told me.

clock_gettime(CLOCK_REALTIME, {1240262824, 400770067}) = 0
sendto(3, "\314\335\356\377\300\250\24p#)\0\0\23\324\207\264\377\377\377\377\0\1\2\3\4\5\6\7\10\t\0\0\3"..., 300, 0, {sa_family=AF_INET, sin_port=htons(9000), sin_addr=inet_addr("192.168.20.112")}, 16) = 300
recvfrom(3, "\314\335\356\377\300\250\24p#)\0\0\23\324\207\264\377\377\377\377\0\1\2\3\4\5\6\7\10\t\0\0\3"..., 300, 0, NULL, NULL) = 300
clock_gettime(CLOCK_REALTIME, {1240262824, 401108324}) = 0
clock_gettime(CLOCK_REALTIME, {1240262824, 401177464}) = 0

cost is about 384 ns on this 32bits 3GHz machine (without strace), while its 51 ns on 64 bit

Anyway I'll try to boot 64bit kernels on those 32bit distro machines in following days...