|
Sorry about the confusion with the language.
Yes, 1 thread = 1 perl process. Yes, 2*100Mb/sec initially.
My test machine had a gig card, through a gig switch to the two 100Mb/s
cards on the servers.
Rest of the response inline.
> -----Original Message-----
> From: mack@xxxxxxxxxxxx [mailto:mack@xxxxxxxxxxxx]On Behalf Of Joseph
> Mack
> Sent: Thursday, September 21, 2000 3:34 PM
> To: Jeffrey A Schoolcraft
> Cc: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
> Subject: Re: Direct Routing sucking up System Resources?
>
>
> Jeffrey A Schoolcraft wrote:
> >
> > I have set up an LVS managing two servers with rr:
> >
> > [root@lb1 /root]# ipvsadm -L -n
> > IP Virtual Server version 0.9.16 (size=16384)
> > Prot LocalAddress:Port Scheduler Flags
> > -> RemoteAddress:Port Forward Weight ActiveConn InActConn
> > TCP 192.168.33.169:80 rr persistent 100
> > -> 192.168.33.113:80 Route 1 0 0
> > -> 192.168.33.114:80 Route 1 0 0
> >
> > ifconfig:
> > eth0:0 Link encap:Ethernet HWaddr 00:D0:B7:1E:8D:56
> > inet addr:192.168.33.169 Bcast:192.168.33.169
> > Mask:255.255.255.255
> > UP BROADCAST RUNNING MULTICAST MTU:1500 Metric:1
> > Interrupt:5 Base address:0xef00 Memory:d0873000-d0873900
> >
> > (169) is a Virtual IP address.
> >
> > When I run two test scripts against it (basically a perl for loop with
> > wget/a bunch of files/) performance is fine,
> > I get 100Mbits/sec from both machines.
>
> is that a total of 2*100Mpbs or 1*100Mbs?
>
> > System resources on the lvs machine
> > are fine with just two threads.
>
> is a "thread" == client perl script process?
>
> When I bump this number up to 4 my net
> > performance drops way down,
>
> does "performance" == throughput?
>
yes, it does.
> and load on the lvs goes up to about 50-60%.
> > (LVS is running all by itself on this server).
>
> where/how do you measure this? I've never got above 5% with
> VS-DR monitoring with top. Then I've never put through more
> than 50Mbps either.
>
I measured it with top.
> > When I bump it up to 6 or
> > more, performance is terrible and the LVS get's killed,
>
> do you mean that the throughput dropped?
yes.
>
> (like 90-100% system
> > resources, as reported by top).
>
> > I thought we should have no overhead running a DR approach.
>
> Shifting 100Mbps of packets requires some work, but I wouldn't have
> expected the results you got.
>
> you sound like you've pushed VS-DR harder than anyone else. I wouldn't
> have expected that the director be working any harder than if it was
> just routing the same number of packets. So I don't know why this
> happened.
cool.
>
> I assume you have 100Mbps ethernet and 1 client. What is the throughput
> returning to the client for each of these tests and how does that compare
> to the throughput you get when getting the files directly from
> the real-servers?
having tried this yet. I imagine throughput would drop. I thought about
adding extra nics into
each machine to see if we could get more performance that way, but it didn't
seem to be the case.
> How does this compare to changing the LVS director to just a router for
> the two realservers.
>
> You must have a healthy set of disks on the real-servers to get 100Mbps
> of files back from them.
>
yeah, healthy, fast, and raid.
> Joe
> --
> Joseph Mack PhD, Senior Systems Engineer, Lockheed Martin
> contractor to the National Environmental Supercomputer Center,
> mailto:mack.joseph@xxxxxxx ph# 919-541-0007, RTP, NC, USA
Jeff
|
