the write strategy of the client implement affects the set performance dramatically?

[Maxim Dounin]

Hello!

On Wed, Mar 05, 2008 at 12:54:44PM +1100, Dean Harding wrote:

> Maxim Dounin wrote:
>> You may note that first packet was ack'ed after 100ms - default delayed 
>> ack timeout in FreeBSD.  Since your next packet is small - OS waits before 
>> sending it, and sends only after previous packet was ack'ed.  And since 
>> memcached don't get the whole request, it has nothing to answer - so ack 
>> was sent only after timeout.
>>
>> Setting net.inet.tcp.delayed_ack to 0 makes times much more real:
>
> I would suggest that you simply send the packet in one write() call, rather 
> than setting TCP_NODELAY. Since memcache has a maximum size of 1MB anyway, 
> you're not going to be creating impossibly big memory buffers (hopefully).

Really, single write() call is wrong way to go. In addition to the 
unneeded buffer allocations this will kill performance due to 
memcpy() involved for whole data chunk.

Writev() is better, but depending on task you may not be able to 
use it. E.g. for 'noreply' commands recently introduced you have 
to use TCP_NODELAY anyway.

Maxim Dounin