FC4 possible kernel 2.6.15 issue

[Skunk Worx]

Skunk Worx wrote:
> Skunk Worx wrote:
>> Under 2.6.14, there was one or two of these 40 ms. calls out of 48,000 
>> calls total. Under 2.6.15, we're seeing about 3,000 40 ms. of these 
>> calls out of 48,000 calls total.
>>

I am posting for completion. This performance hit was due to the new 
2.6.15 "ABC" (Appropriate Byte Count) algorithm. There is an IETF RFC 
for this, RFC3465.

The purpose is to decrease internet congestion, but it can have a 
performance impact on poorly written network code (for example setting 
TCP_NODELAY on sockets and sending packets with small payloads)

It is possible to disable this feature with sysctl of net.ipv4.tcp_abc.

In our case, all is well...we found a setting in the GSOAP toolkit that 
restores the original performance...we're FC5 ready!

--
SW