LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

RE: lvs bottlekneck

To: "'Drew Streib '" <ds@xxxxxxxxxxx>, "'Cono D'Elia '" <conod@xxxxxxxx>, "'lvs-users@xxxxxxxxxxxxxxxxxxxxxx '" <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: RE: lvs bottlekneck
From: Dan <dan@xxxxxxxxxxx>
Date: Fri, 12 May 2000 13:05:45 -0700
 I have already run into director bottlenecks under normal traffic. They are
present because I'm using NAT, sustain a large number of connections, and
create outbound connections for many of my inbound connections (read:
proxy). The problem is that the Linux kernel only supports 4096 masquerading
(NAT) connections outbound. Going beyond that brings the system to its
knees. I had to modify the kernel & recompile to up this number. Also, in
regard to sufficiency of memory - 64M is not enough for my app. I need to
sustain 11000 simulataneous connections which amounts to a hash table of
2^19. That's 68M of RAM just for the hash table. I realize that many apps
are not this severe, but they are points to keep in mind when you start
pushing past the base 2^12 table size.

-d


-----Original Message-----
From: Drew Streib
To: Cono D'Elia; lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Sent: 5/12/00 11:30 AM
Subject: Re: lvs bottlekneck

On Sat, May 13, 2000 at 02:16:02PM -0700, Cono D'Elia wrote:
> Hello,
> 
> What do you do when you reach the point where a single linux load
balancer box becomes the bottlekneck? What about configuring the DNS
entry for the www site to have multiple ips ( one for each linux load
balancer box) and to be round robin... ie...

I want to see the application in which a fast Intel based LVS director
becomes
the bottleneck.

I'll be performing "official" benchmarks on this in the near future, and
from what I'm seeing now, I don't expect anything less than 200-300Mbit
minimum bandwidth using NAT modes for a fast Intel box. Direct Routing
would
be even faster.

There are < 10 sites in the world which sustain this much bandwidth.

Having a second box for failover is a good idea... but I'd question 
a suspected lvs bottleneck.

Has anyone on this list actually benchmarked under real conditions
and found the lvs director was actually slowing traffic?

-drew

-- 
-------
Drew Streib <d@xxxxxxxxxxx> 408.542.5725

Tech Marketing/Benchmarking, VA Linux Systems     | <dtype@xxxxxxxxxxx>
Sr Developer, Community Liason, SourceForge       |
<dtype@xxxxxxxxxxxxxxx>
System Administrator, Linux International         | <dtype@xxxxxx>
Admirer, Occasional Programmer, Linux.com         | <dtype@xxxxxxxxx>



<Prev in Thread] Current Thread [Next in Thread>