[c-nsp] Performance issue on link

Tony td_miles at yahoo.com
Mon Apr 1 22:54:06 EDT 2013 

Yes, if it's possible, bypass the routers and test directly through the switches. This will rule out any shaping/policing issues being done on the router.

UDP has no flow control and only attempts to send traffic at the speed you tell it to (ie. 40Mbps) so it won't exceed the bandwidth of the link. TCP tries to send as fast as it can (ie. 100Mbps or Gbps on gig NIC/port) and when packets get dropped it implements retries, windowing & flow control to achieve the best it can.

It could also be the carrier policing extremely hard, but they shouldn't really.


regards,
Tony.






>________________________________
> From: CiscoNSP List <cisconsp_list at hotmail.com>
>To: "td_miles at yahoo.com" <td_miles at yahoo.com>; Azher Mughal <azher at hep.caltech.edu> 
>Cc: "cisco-nsp at puck.nether.net" <cisco-nsp at puck.nether.net> 
>Sent: Tuesday, 2 April 2013 12:07 PM
>Subject: RE: [c-nsp] Performance issue on link
> 
>
> 
>
>Hi Tony - Link is new, no duplex issues (On our side, nor reported by carrier)
>
>
>
>
>No traffic on the link.
>
>
>
>
>udp test (both directions) looks fine (Hopefully formatting isnt screwed up - but achieve ~38.5Mb/sec both directions):
>
>
>
>
>
>
>[ ID] Interval       Transfer     Bandwidth
>
>
>[  5]  0.0-10.0 sec  47.7 MBytes  40.0 Mbits/sec
>
>
>[  5] Sent 34015 datagrams
>
>
>[  3]  0.0-10.0 sec  46.0 MBytes  38.6 Mbits/sec  0.040 ms 1180/34014 (3.5%)
>
>
>[  3]  0.0-10.0 sec  1 datagrams received out-of-order
>
>
>[  5] Server Report:
>
>
>[  5]  0.0-10.0 sec  46.0 MBytes  38.5 Mbits/sec  0.051 ms 1233/34014 (3.6%)
>
>
>[  5]  0.0-10.0 sec  1 datagrams received out-of-order
>
>
>
>
>
>
>The 2 linux servers are on separate L3 subnets (Connected to the local router in each POP) - I can potentially trunk a new vlan between the 2 POPs and have them connected without touching the routers...UDP tests are looking good though, so if it was an issue with the router(s), wouldn't I see problems with UDP also?
>
>
>
>
>Thanks again for your assistance, much appreciated.
>
>
>
>
>
>
>________________________________
>Date: Mon, 1 Apr 2013 18:19:40 -0700
>From: td_miles at yahoo.com
>Subject: Re: [c-nsp] Performance issue on link
>To: cisconsp_list at hotmail.com; azher at hep.caltech.edu
>CC: cisco-nsp at puck.nether.net
>
>
>That doesn't sound at all good. No way should you need that many (25) sessions.
>
>Is there any other traffic on the link ? Is it brand new ? Any duplex mismatch type issues ?
>
>I would suggest running an "iperf -u -d" to run UDP in both directions at the same time, might turn up something unusual. Remove any other devices (ie. your routers) before testing so you will know that is not a problem. I assume your two linux test machines are on the same L3 subnet from what you've described ?
>
>
>regards,
>Tony.
>
>
>
>
>
>
>
>
>>________________________________
>> From: CiscoNSP List <cisconsp_list at hotmail.com>
>>To: "td_miles at yahoo.com" <td_miles at yahoo.com>; Azher Mughal <azher at hep.caltech.edu> 
>>Cc: "cisco-nsp at puck.nether.net" <cisco-nsp at puck.nether.net> 
>>Sent: Tuesday, 2 April 2013 10:37 AM
>>Subject: RE: [c-nsp] Performance issue on link
>> 
>>
>> 
>>Sorry - Forgot to add:
>>
>>
>>
>>
>>POP A -> POP B - only requires 8 sessions, but POP B -> POP A requires 25 sessions to achieve ~38Mb?
>>
>>
>>
>>
>>
>>
>>
>>> From: cisconsp_list at hotmail.com
>>> To: td_miles at yahoo.com; azher at hep.caltech.edu
>>> Date: Tue, 2 Apr 2013 11:28:50 +1100
>>> CC: cisco-nsp at puck.nether.net
>>> Subject: Re: [c-nsp] Performance issue on link
>>> 
>>> Thanks Tony - With multiple sessions (8), I can achieve ~38Mb/sec consistently.
>>> 
>>> Is it simply not feasible to expect ~40Mb with a single tcp transfer (Without significant adjustments to both Linux servers?)
>>> 
>>> 
>>> > Date: Mon, 1 Apr 2013 17:10:42 -0700
>>> > From: td_miles at yahoo.com
>>> > Subject: Re: [c-nsp] Performance issue on link
>>> > To: cisconsp_list at hotmail.com;
 azher at hep.caltech.edu
>>> > CC: cisco-nsp at puck.nether.net
>>> > 
>>> > Try using multiple TCP sessions  "iperf -P 5"  (note - capital "P").
>>> > 
>>> > 
>>> > regards,
>>> > Tony.
>>> > 
>>> > 
>>> > 
>>> > 
>>> > 
>>> > ----- Original Message -----
>>> > > From: CiscoNSP List <cisconsp_list at hotmail.com>
>>> > > To: Azher Mughal <azher at hep.caltech.edu>
>>> > > Cc: "cisco-nsp at puck.nether.net" <cisco-nsp at puck.nether.net>
>>> > > Sent: Tuesday, 2 April 2013 9:51 AM
>>> > > Subject: Re: [c-nsp] Performance issue on link
>>> > > 
>>> > >T hanks - Ill check them out now
>>> > > 
>>> > >>  Date: Mon, 1 Apr 2013 16:16:36 -0700
>>> > >>  From: azher at hep.caltech.edu
>>> > >>  To: cisconsp_list at hotmail.com
>>> > >>  CC:
 cisco-nsp at puck.nether.net
>>> > >>  Subject: Re: [c-nsp] Performance issue on link
>>> > >> 
>>> > >>  These might be helpful:
>>> > >> 
>>> > >>  http://fasterdata.es.net/network-tuning/
>>> > >> 
>>> > >> 
>>> > > http://www.mellanox.com/related-docs/prod_software/Performance_Tuning_Guide_for_Mellanox_Network_Adapters.pdf
>>> > >> 
>>> > >>  -Azher
>>> > >> 
>>> > >>  On 4/1/2013 3:51 PM, CiscoNSP List wrote:
>>> > >>  > Hi,
>>> > >>  >
>>> > >>  > We have a 40Mb link between 2 POPs - Latency ~65m/sec (No packet-loss)
>>> > >>  >
>>> > >>  > POP A Is a 7301 and 2960POP B is a 7200 and 4948
>>> > >>  >
>>> > >>  > 40Mb link is connected to the two switches (L2), and then a trunk link 
>>>
 > > to both routers for all L3.
>>> > >>  >
>>> > >>  > Have a Linux server connected to both switches, and achieve the 
>>> > > following performance:
>>> > >>  >
>>> > >>  > IPERF (UDP)
>>> > >>  > POP A -> POP B - 38.5Mb/secPOP B -> POP A - 38.5Mb/sec
>>> > >>  >
>>> > >>  > IPERF (TCP)
>>> > >>  > POP A -> POP B - ~20Mb/secPOP B -> POP A - ~12Mb/sec
>>> > >>  >
>>> > >>  > FTP
>>> > >>  > POP A -> POP B - ~38Mb/secPOP B -> POP A - ~16Mb/sec
>>> > >>  >
>>> > >>  > WGET
>>> > >>  > POP A -> POP B - ~30Mb/sec POP B -> POP A - ~16Mb/sec
>>> > >>  >
>>> > >>  > Any suggestions on why I am seeing poor performance with TCP 
>>> > > transfers? (Especially
 POP B -> POP A direction) - I've tried adjusting 
>>> > > the window size in IPERF but it actually made the results worse?
>>> > >>  > Thanks in advance.
>>> > >>  >
>>> 
>>> _______________________________________________
>>> cisco-nsp mailing list  cisco-nsp at puck.nether.net
>>> https://puck.nether.net/mailman/listinfo/cisco-nsp
>>> archive at http://puck.nether.net/pipermail/cisco-nsp/
>>
>>
>>
>
>

More information about the cisco-nsp mailing list