[Spread-users] [info] a change in events.c to use /dev/poll on solaris 8

[Jonathan Stanton]

That's great! Customizing the Spread event handling for system specific
interfaces that offer better performance (like /dev/poll and FreeBSD's
kevents) is something I've been interested in for awhile.

There is a relationshoip between token messages and user messages. Each
time the token rotates around the daemons, the flow control regulates how
many "packets" each daemon can send. If each user message is larger then a
packet (1400 bytes roughly) then it is fragmented and requires multiple
packets to be sent per user message. If the user message is smaller then a
packet and several user messages are sent quickly by a set of locally
connected clients, the daemon will try to pack multiple user messages into
one network packet. This increases message throughput noticably for small
messages. 

Tuning the flow control parameters on the token can definitely improve
performance if you have known patterns of loss and sending.

Jonathan

On Sat, Dec 21, 2002 at 04:02:01PM +0900, ¿øÅÂȯ wrote:
> Hi,
> 
> I changed spread-3.17 to run on 64-bit solaris 8 with **/dev/poll**.
> Most of changes were made in events.c/session.c
> 
> As a result, I got the following pros/cons:
> 	pros) performance benefits : about 30% or higher performance up
> 		in terms of msg/sec. (As a test tool, I used spflooder/my own echo tool)
> 		When I started this work, I expected to lower cpu load with the same
> 		traffic value.
> 	cons) Loss of benefits of event priority 
> 		But, I think if I put more time in the modification, I would be
> 		able to handle it. ;-)
> 
> This result comes only from solaris 8, not for every other platform.
> 
> I have one question. Is there any relationship between # of user messages and
> # of token-related messages ?
> I think if LAN is very clean, it's possible to reduce the token-related messages.
> 
> The reason I don't attach my change to this mail, I don't want to confuse
> spread-users and some part of works was not completed yet.
> Anyway, I hope to this experiment is useful to spread-users.
> 
> 
> -----------------
> Tawan Won
> 
> _______________________________________________
> Spread-users mailing list
> Spread-users at lists.spread.org
> http://lists.spread.org/mailman/listinfo/spread-users

-- 
-------------------------------------------------------
Jonathan R. Stanton         jonathan at cs.jhu.edu
Dept. of Computer Science   
Johns Hopkins University    
-------------------------------------------------------