latency doubled on tg3 device from 2.6.11 to 2.6.12

latency doubled on tg3 device from 2.6.11 to 2.6.12

[Rick Warner]

Hello,
 We have been testing latency and bandwidth using our proprietary MPI link 
checker tool (http://www.microway.com/mpilinkchecker.html) and have found 
that the latency increased from ~25ms to ~45ms going from 2.6.11 to 2.6.12.  
2.6.13 has the same result.  We also tried the latest bcm5700 from broadcom 
(8.2.18) and got the same ~45ms latencies.  This was tried on several 
different opteron and em64t motherboards.

 We see 20-25ms latencies with the e1000 driver (with some module options) on 
all 3 kernel versions.  For those interested, the e1000 options used are:

 InterruptThrottleRate=0 RxIntDelay=0 TxIntDelay=0 RxAbsIntDelay=0 
TxAbsIntDelay=0

 Digging through source, it seems that a new locking mechanism for tg3 was put 
in place in 2.6.12.  Is this the likely cause?  What can we do to restore our 
lower latency?


-- 
Richard Warner
Lead Systems Integrator
Microway, Inc
(508)732-5517

Re: latency doubled on tg3 device from 2.6.11 to 2.6.12

[Eric Dumazet]

Rick Warner a écrit :
> Hello,
>  We have been testing latency and bandwidth using our proprietary MPI link 
> checker tool (http://www.microway.com/mpilinkchecker.html) and have found 
> that the latency increased from ~25ms to ~45ms going from 2.6.11 to 2.6.12.  
> 2.6.13 has the same result.  We also tried the latest bcm5700 from broadcom 
> (8.2.18) and got the same ~45ms latencies.  This was tried on several 
> different opteron and em64t motherboards.
> 
>  We see 20-25ms latencies with the e1000 driver (with some module options) on 
> all 3 kernel versions.  For those interested, the e1000 options used are:
> 
>  InterruptThrottleRate=0 RxIntDelay=0 TxIntDelay=0 RxAbsIntDelay=0 
> TxAbsIntDelay=0
> 
>  Digging through source, it seems that a new locking mechanism for tg3 was put 
> in place in 2.6.12.  Is this the likely cause?  What can we do to restore our 
> lower latency?
> 
> 

Could you please define latency ?

tg3 driver was recently updated to use coalescing.

So when the nic receives one frame, it may delay up to XXXX us ( XXXX < 1024) 
the interrupt.

But 25 ms is far more than 1024 us, so I dont think this coalescing can 
explain your problem.

The HZ change from 1000 to 250 could be the root of the problem ?

Using a simple ping between 2 machines with tg3, I get less than 1ms time.

Eric

Re: latency doubled on tg3 device from 2.6.11 to 2.6.12

[Rick Warner]

On Thursday 01 September 2005 05:46 pm, Eric Dumazet wrote:
> Rick Warner a écrit :
> > Hello,
> >  We have been testing latency and bandwidth using our proprietary MPI
> > link checker tool (http://www.microway.com/mpilinkchecker.html) and have
> > found that the latency increased from ~25ms to ~45ms going from 2.6.11 to
> > 2.6.12. 2.6.13 has the same result.  We also tried the latest bcm5700
> > from broadcom (8.2.18) and got the same ~45ms latencies.  This was tried
> > on several different opteron and em64t motherboards.
> >
> >  We see 20-25ms latencies with the e1000 driver (with some module
> > options) on all 3 kernel versions.  For those interested, the e1000
> > options used are:
> >
> >  InterruptThrottleRate=0 RxIntDelay=0 TxIntDelay=0 RxAbsIntDelay=0
> > TxAbsIntDelay=0
> >
> >  Digging through source, it seems that a new locking mechanism for tg3
> > was put in place in 2.6.12.  Is this the likely cause?  What can we do to
> > restore our lower latency?
>
> Could you please define latency ?
>
> tg3 driver was recently updated to use coalescing.
>
> So when the nic receives one frame, it may delay up to XXXX us ( XXXX <
> 1024) the interrupt.
>
> But 25 ms is far more than 1024 us, so I dont think this coalescing can
> explain your problem.
>
> The HZ change from 1000 to 250 could be the root of the problem ?
>
> Using a simple ping between 2 machines with tg3, I get less than 1ms time.
>
> Eric

Our mpi link checker tool does a 1 way transfer of a 0 byte message (+ header) 
and times it to each system (in addition to other tests).  All systems in a 
cluster are showing the same high latency.  The numbers I gave were supposed 
to be us, I used the wrong unit by accident.

Low latency is often essential to clustered applications.  While a delay of up 
to 1024 us doesn't affect regular communications too much, it may severely 
affect mpi jobs.

Thanks.

-- 
Richard Warner
Lead Systems Integrator
Microway, Inc
(508)732-5517

Re: latency doubled on tg3 device from 2.6.11 to 2.6.12

[Eric Dumazet]

Rick Warner a écrit :
> On Thursday 01 September 2005 05:46 pm, Eric Dumazet wrote:
> 
>>Rick Warner a écrit :
>>
>>>Hello,
>>> We have been testing latency and bandwidth using our proprietary MPI
>>>link checker tool (http://www.microway.com/mpilinkchecker.html) and have
>>>found that the latency increased from ~25ms to ~45ms going from 2.6.11 to
>>>2.6.12. 2.6.13 has the same result.  We also tried the latest bcm5700
>>>from broadcom (8.2.18) and got the same ~45ms latencies.  This was tried
>>>on several different opteron and em64t motherboards.
>>>
>>> We see 20-25ms latencies with the e1000 driver (with some module
>>>options) on all 3 kernel versions.  For those interested, the e1000
>>>options used are:
>>>
>>> InterruptThrottleRate=0 RxIntDelay=0 TxIntDelay=0 RxAbsIntDelay=0
>>>TxAbsIntDelay=0
>>>
>>> Digging through source, it seems that a new locking mechanism for tg3
>>>was put in place in 2.6.12.  Is this the likely cause?  What can we do to
>>>restore our lower latency?
>>
>>Could you please define latency ?
>>
>>tg3 driver was recently updated to use coalescing.
>>
>>So when the nic receives one frame, it may delay up to XXXX us ( XXXX <
>>1024) the interrupt.
>>
>>But 25 ms is far more than 1024 us, so I dont think this coalescing can
>>explain your problem.
>>
>>The HZ change from 1000 to 250 could be the root of the problem ?
>>
>>Using a simple ping between 2 machines with tg3, I get less than 1ms time.
>>
>>Eric
> 
> 
> Our mpi link checker tool does a 1 way transfer of a 0 byte message (+ header) 
> and times it to each system (in addition to other tests).  All systems in a 
> cluster are showing the same high latency.  The numbers I gave were supposed 
> to be us, I used the wrong unit by accident.
> 
> Low latency is often essential to clustered applications.  While a delay of up 
> to 1024 us doesn't affect regular communications too much, it may severely 
> affect mpi jobs.
> 
> Thanks.
> 

OK

tg3 driver set a rx-usecs of 20us per default.
This is certainly the root of your problem.

Try to lower it on mpi links ?

ethtool -C eth0 rx-usecs 0

Eric