Mailing List Archive

On Tue, Dec 26, 2000 at 10:37:08AM -0800, Florin Andrei wrote:
> I did some research on the mailing list archive, but i
> was unable to find a match for this problem...
>
> I understand that system latency is a critical
> parameter for LVS. So, a fast network interface card
> (NIC) might be an advantage for the balancer box.
>
> What NICs do you reccomend for using in a LVS
> balancer, on Intel architecture?

I don't believe that anyone has done an extensive analysis of the
performance of different network cards with regards to LVS. In terms of
100Mb/s it would make sense to me that you are going to want a reasonable
quality card. Intel eepro, DEC tulip based and 3Com cards fall into this
category for better or for worse. I've done all my work with eepro crads
because that's what I can get my hands on.

In terms of 1000Mb/s I've been having good success with Netgear acenic
cards of late, but again thats because I can get my hands on them.

To be honest, I think that the bigest problem you are going to come up
against is bus speed. Some performace testing I've been privy to lately
indicates a big advantage in using PCI 64bit/66MHz over 32bit/33MHz,
though this only really comes into play with gigabit.


> Is there an FAQ for LVS? I think this question is a
> typical FAQ entry...

No, but there is a HOWTO and a mailing list archive.

--
Horms

3com and Intel cards are a good choice. I am using both in production.

See their web sites for the latest Linux driver modules which you can download, compile and load into your system (on Red Hat via /etc/conf.modules for example).

--K >From: Florin Andrei >Reply-To: lvs-users@LinuxVirtualServer.org >To: LVS users >Subject: reccomended NIC >Date: Tue, 26 Dec 2000 10:37:08 -0800 (PST) > >I did some research on the mailing list archive, but i >was unable to find a match for this problem... > >I understand that system latency is a critical >parameter for LVS. So, a fast network interface card >(NIC) might be an advantage for the balancer box. > >What NICs do you reccomend for using in a LVS >balancer, on Intel architecture? > >Is there an FAQ for LVS? I think this question is a >typical FAQ entry... > >-- >Florin Andrei > > >__________________________________________________ >Do You Yahoo!? >Yahoo! Shopping - Thousands of Stores. Millions of Products. >http://shopping.yahoo.com/ > >_______________________________________________ >LinuxVirtualServer.org mailing list - lvs-users@LinuxVirtualServer.org >Send requests to lvs-users-request@LinuxVirtualServer.org >or go to http://www.in-addr.de/mailman/listinfo/lvs-users

---

Get your FREE download of MSN Explorer at http://explorer.msn.com"]http://explorer.msn.com

I use Intel cards, since 3Com cards had once locked up
the LVS box during stress tests. Never had that problem
with Intel cards.

At 08:20 PM 12/26/00 +0000, Lorn Kay wrote:


>3com and Intel cards are a good choice. I am using both in production.
>
>See their web sites for the latest Linux driver modules which you can download, compile and load into your system (on Red Hat via /etc/conf.modules for example).
>
>--K
>>From: Florin Andrei
>>Reply-To: lvs-users@LinuxVirtualServer.org
>>To: LVS users
>>Subject: reccomended NIC
>>Date: Tue, 26 Dec 2000 10:37:08 -0800 (PST)
>>
>>I did some research on the mailing list archive, but i
>>was unable to find a match for this problem...
>>
>>I understand that system latency is a critical
>>parameter for LVS. So, a fast network interface card
>>(NIC) might be an advantage for the balancer box.
>>
>>What NICs do you reccomend for using in a LVS
>>balancer, on Intel architecture?
>>
>>Is there an FAQ for LVS? I think this question is a
>>typical FAQ entry...
>>
>>--
>>Florin Andrei
>>
>>
>>__________________________________________________
>>Do You Yahoo!?
>>Yahoo! Shopping - Thousands of Stores. Millions of Products.
>>http://shopping.yahoo.com/
>>
>>_______________________________________________
>>LinuxVirtualServer.org mailing list - lvs-users@LinuxVirtualServer.org
>>Send requests to lvs-users-request@LinuxVirtualServer.org
>>or go to http://www.in-addr.de/mailman/listinfo/lvs-users
>
>
>----------
>Get your FREE download of MSN Explorer at <http://explorer.msn.com>http://explorer.msn.com
>_______________________________________________ LinuxVirtualServer.org mailing list - lvs-users@LinuxVirtualServer.org Send requests to lvs-users-request@LinuxVirtualServer.org or go to http://www.in-addr.de/mailman/listinfo/lvs-users

Hi,

> I understand that system latency is a critical
> parameter for LVS. So, a fast network interface card
> (NIC) might be an advantage for the balancer box.

Depending on what you do, LVS-DR or LVS-NAT, have a look at
Joe's stress tests hosted on the LVs-website. At a certain
point I could imagine that also the PCI-bus will be a limiting
factor. At least when you're building a load balancer with
more then 4 NIC's that tries to handle 60Mbit/s requests on
each NIC (although you need quite a fat backbone)

> What NICs do you reccomend for using in a LVS
> balancer, on Intel architecture?

Mhh, if you refer to the CPU with Intel architecture, Horms
already gave you a reasonable answer, however if you refer to
the mainboard I might just add that eepro100 cards perform
very good with Intel Boards, especially the L440GX+ (monsterboards)
But as experience shows it's more the drivers status that's
limiting. The standard cards (like 3com, DEC and eepro100) all
have similar architecture and builtin intelligence, however
the driver doesn't always take advantage of their capabilities
therefore you might see higher troughput on test with f.e DEC
cards on NT then on Linux because the NT driver does take advantage
of the hardware TCP sanity checks and stuff like that and under
Linux not because of the compatibility with the rest of the TCP/
IP stack implementation. See, we're talking about some 1-2Mbits/s
of requests of difference and if you really would experience a
bottleneck with some of the cards Horms mentioned, I'll gladly
help you speeding them up by tweaking some bits ;)

> Is there an FAQ for LVS? I think this question is a
> typical FAQ entry...

I was going to suggest it to Joe.

regards,
Roberto Nibali, ratz

--
mailto: `echo NrOatSz@tPacA.cMh | sed 's/[NOSPAM]//g'`

On Tue, 2 Jan 2001, ratz wrote:

> > Is there an FAQ for LVS? I think this question is a
> > typical FAQ entry...
>
> I was going to suggest it to Joe.

we don't have any hard data yet. I can't imagine anyone crazy enough to
sit down with a bucket full of cards and a room full of mobos with
different chip sets and different kernels and find out the real story any
time real soon either. I would say that if your setup passes the netpipe
test, then you are done.

My performance tests show that the tcp race condition in 2.2.x has a
greater effect than the cards. I don;'t know if this is a problem in
2.4.x.

I also found that the Netgear cards didn't work on one particular set of
ancient hardware that the eepro100 did fine with. I don't think this is a
real indictment of the Netgear cards though as they work fine with other
mobos.

Joe

--
Joseph Mack mack@ncifcrf.gov

Joseph Mack wrote:
> we don't have any hard data yet. I can't imagine anyone crazy enough to
> sit down with a bucket full of cards and a room full of mobos with
> different chip sets and different kernels and find out the real story any
> time real soon either. I would say that if your setup passes the netpipe
> test, then you are done.

german magazine c't made a test last year with various 10/100/1000 mbit
cards under linux, windows and os/2. I'll look through my archieves if I
find it. I think I could scan some nice graphs so we can watch at them.
AFAIR 3com and SysKonnekt made the race....anyway, I'll take a look and
report you soon.


--

Regards,

Wiktor Wodecki
Unix-Administration
Wapme-Systems Inc.

On Wed, 10 Jan 2001, Wiktor Wodecki wrote:

> Joseph Mack wrote:
> > we don't have any hard data yet. I can't imagine anyone crazy enough to
> > sit down with a bucket full of cards and a room full of mobos with
> > different chip sets and different kernels and find out the real story any
> > time real soon either.

> german magazine c't made a test last year with various 10/100/1000 mbit
> cards under linux, windows and os/2.

so much for my imagination :)

I'll look through my archieves if I
> find it. I think I could scan some nice graphs so we can watch at them.
> AFAIR 3com and SysKonnekt made the race....anyway, I'll take a look and
> report you soon.

thanks

Joe

--
Joseph Mack mack@ncifcrf.gov