Andrew Gallatin wrote:
Jeff Trawick wrote:
Server CPU use:
user space CPU is pretty close, with sendfile using the least (23.64
sec, 24.65, 24.74)
system space CPU reduction with sendfile is significant (52.55 sec,
69.25 sec, 68.84 sec)
How did you measure system space CPU reduction? A utility like "time"
that measures overhead only in the process context of the process you're
measuring, or globally, using a tool like vmstat?
definitely process context only
(either the result of Apache calling times() or ps I imagine)
The problem is that due to what I consider to be hacks to work around
broken drivers, packets sent by sendfile are queued up freed later
from a different context. This is done without regard to CPU
affinity, etc, and causes excessive cache misses.
A tool like "time" will not show this extra overhead, but a tool
like vmstat which measures system-wide overhead will show the
extra load caused by the packets being freed in a different context.
This seems worst on amd64 with 10GbE networks. It may not be noticeable
with 1G networks.
interesting
_______________________________________________
networking-discuss mailing list
[email protected]
