Velocity Reviews - Computer Hardware Reviews

Velocity Reviews > Newsgroups > Computing > Cisco > Throughput of your Gigabit switch? (performance problems with 3508G-XL)

Reply
Thread Tools

Throughput of your Gigabit switch? (performance problems with 3508G-XL)

 
 
Thomas Arens
Guest
Posts: n/a
 
      10-04-2004
http://www.velocityreviews.com/forums/(E-Mail Removed) schrieb:
>
> Also keep in mind that the 3508 doesn't really support full gigabit
> speeds on all ports at once. Ports 1&2, 3&4, 5&6, and 7&8 share an
> asic (4 total). Each asic can only support 1 gigabit throughput.


Thanks for the information, but I think in this benchmark that's
irrelevant, since one node sends and the other receives.

It is interesting thinking of the MPI performance with 8 nodes though.
 
Reply With Quote
 
 
 
 
Thomas Arens
Guest
Posts: n/a
 
      10-04-2004
Ivan Ostreš wrote:
>
> Actually it would be pretty stupid to test maximal troughput using TCP
> anyway. If you really want to get maximum you should use some UDP based
> tool.


I don't think so, because TCP is quite realistic for many applications.
Benchmarking UDP performance would deliver quite theoretical results.
 
Reply With Quote
 
 
 
 
Thomas Arens
Guest
Posts: n/a
 
      10-04-2004
Martin Bilgrav schrieb:

> I seem to recall that the 3500 have limited framesize support, if any-
> beyond 2k frames.


Jumbo frames are not supported, yes.

> I would look into that, If I were you.
> Looks like the switch are config'ed for some MTU and/or frame-size, that has
> a mis-match between the NIC and the port.


I checked the MTU sizes. They should be correct.

> (btw was that copper ports in the
> GBIC?)


Yes.

> btw you dont need cross-over for 1000base-t NICs, and remember that
> 1000base-t uses all 4 pairs as well, so be very sure you use brand new Cat-6
> or 5e (I use Cat.6 with each pair seperated by a plastic X inside the
> cable).


I used quite short (a few meters long) Cat 5e cables with the
inscription "verified for gigabit ethernet" or something like that. I do
not have Cat 6 cables to play with...

> But I would use an other switch that supports large frames.


Hehe... you know, money _is_ an issue.

Thanks!

Thomas Arens
Chair for Operating Systems (http://www.lfbs.rwth-aachen.de)
RWTH Aachen University (http://www.rwth-aachen.de)
 
Reply With Quote
 
Thomas Arens
Guest
Posts: n/a
 
      10-04-2004
Chris Thomas schrieb:

> at a first
> approximation, I didn't see any difference between going through the
> switch and using a crossover
> [...]
> I would agree that your 'switch' numbers are really poor, but
> have no suggestions beyond that.


Thanks, that's good to know.
 
Reply With Quote
 
Ivan Ostreš
Guest
Posts: n/a
 
      10-04-2004
In article <(E-Mail Removed)>, http://www.velocityreviews.com/forums/(E-Mail Removed)-aachen.de
says...
> Ivan Ostreš wrote:
> >
> > Actually it would be pretty stupid to test maximal troughput using TCP
> > anyway. If you really want to get maximum you should use some UDP based
> > tool.

>
> I don't think so, because TCP is quite realistic for many applications.
> Benchmarking UDP performance would deliver quite theoretical results.
>


Just partially true. TCP is quite realistic for many applications but
you will not actually measure the limits of the switch. So, you have to
ask yourself what do you actually want to measure: how fast your
application will be using specified hardware or how much packets
(bytes/whatever) can hardware handle?

It's simple as that. Someone said (i think it was my university teacher)
that measuring something when you haven't defined what do you really
want and didn't choose right way is worth nothing.


--
-Ivan.

*** Use Rot13 to see my eMail address ***
 
Reply With Quote
 
Hansang Bae
Guest
Posts: n/a
 
      10-05-2004
In article <(E-Mail Removed)>, (E-Mail Removed)-aachen.de
says...
> Hello everybody,
>
> I am currently benchmarking several MPI implementations for Windows, and
> I noticed that our Cisco 3508G-XL Gigabit switch (eight 1000Base-X
> ports, 10-Gbps switching fabric) performs quite bad compared to a direct
> connection between two nodes via crossover cable.
>
> I used two nodes (dual Xeon machines, Intel Pro/1000 MT Server Adapters,
> Windows XP Prof. SP1) and the tool "netio" (version 1.14,
> client/server-based, measures TCP/IP-throughput in one direction, Google
> for download ), once connected through the Cisco switch and once
> with a crossover cable.
>
> The results:
>
> Switch:
>
> Packet size 1 KByte: 32580 KByte/s
> Packet size 2 KByte: 37728 KByte/s
> Packet size 4 KByte: 30869 KByte/s
> Packet size 8 KByte: 40347 KByte/s
> Packet size 16 KByte: 69608 KByte/s
> Packet size 32 KByte: 79617 KByte/s
>
> Crossover cable:
>
> Packet size 1 KByte: 74438 KByte/s
> Packet size 2 KByte: 86279 KByte/s
> Packet size 4 KByte: 94309 KByte/s
> Packet size 8 KByte: 105423 KByte/s
> Packet size 16 KByte: 115338 KByte/s
> Packet size 32 KByte: 115516 KByte/s
>
> As you can see, the performance of the switch is really poor. I went
> through the configuration of the switch many times, but I could not find
> anything that would explain the bad performance, the configuration seems
> to be correct.
>
>
> How are your experiences with Gigabit switches regarding performance?
> I would greatly appreciate any statements on this topic.
>
> Many thanks for your time!



Somethings to consider. Take Cisco's claimed numbers and then chop it in
half. That's a start. Do you have the spec sheet handy?

Try the following:
1) use auto/auto on both NICs.
2) Use ports 1/2 instead of 1/7 for example. It could be asic level
issues going on here.



--

hsb

"Somehow I imagined this experience would be more rewarding" Calvin
*************** USE ROT13 TO SEE MY EMAIL ADDRESS ****************
************************************************** ******************
Due to the volume of email that I receive, I may not not be able to
reply to emails sent to my account. Please post a followup instead.
************************************************** ******************
 
Reply With Quote
 
Thomas Arens
Guest
Posts: n/a
 
      10-06-2004
Ivan Ostreš schrieb:
> Just partially true. TCP is quite realistic for many applications but
> you will not actually measure the limits of the switch. So, you have to
> ask yourself what do you actually want to measure: how fast your
> application will be using specified hardware or how much packets
> (bytes/whatever) can hardware handle?


True. And I want to know how fast the switch can handle TCP packets. I
did think about that.

- The background is: I'm benchmarking several MPI implementations for
Windows, and I noticed that the switch seems to be quite slow. Now I'm
trying to find out if the performance is "normal" for a Gigabit switch
and if not, how I could make it perform better.


Best regards,
Thomas Arens
Chair for Operating Systems (http://www.lfbs.rwth-aachen.de)
RWTH Aachen University (http://www.rwth-aachen.de)
 
Reply With Quote
 
Steinar Haug
Guest
Posts: n/a
 
      10-06-2004
[Thomas Arens]

| > Just partially true. TCP is quite realistic for many applications but
| > you will not actually measure the limits of the switch. So, you have to
| > ask yourself what do you actually want to measure: how fast your
| > application will be using specified hardware or how much packets
| > (bytes/whatever) can hardware handle?
|
| True. And I want to know how fast the switch can handle TCP packets. I
| did think about that.

If this is a normal L2 switch is has *no idea* that your packets are TCP
packets.

If you want a realistic measurement of switch throughput, you need
measurement equipment designed for this - Smartbits or similar.

Steinar Haug, Nethelp consulting, (E-Mail Removed)
 
Reply With Quote
 
Ivan Ostreš
Guest
Posts: n/a
 
      10-06-2004
In article <(E-Mail Removed)>, (E-Mail Removed) says...
> [Thomas Arens]
>
> | > Just partially true. TCP is quite realistic for many applications but
> | > you will not actually measure the limits of the switch. So, you have to
> | > ask yourself what do you actually want to measure: how fast your
> | > application will be using specified hardware or how much packets
> | > (bytes/whatever) can hardware handle?
> |
> | True. And I want to know how fast the switch can handle TCP packets. I
> | did think about that.
>
> If this is a normal L2 switch is has *no idea* that your packets are TCP
> packets.
>


Of course it has *no idea* that packets are TCP, but endpoints
(computers) are very aware of that .


--
-Ivan.

*** Use Rot13 to see my eMail address ***
 
Reply With Quote
 
 
 
Reply

Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are Off


Similar Threads
Thread Thread Starter Forum Replies Last Post
Gigabit Throughput yashpal.chouhan@gmail.com Cisco 2 11-23-2006 07:41 AM
Using a Gigabit copper to Gigabit fiber media converter for an IDF rjmnyc@optonline.net Cisco 5 05-06-2006 12:51 PM
catalyst 2948G throughput performance problems ? Martin Bilgrav Cisco 1 12-30-2003 01:51 AM
PIX 506: what does "Throughput: Limited" mean? Malcolm Dunnett Cisco 0 10-18-2003 04:30 PM
any VHDL books on low resource(throughput, area, power trade-off) design? walala VHDL 0 09-08-2003 12:07 AM



Advertisments