Xen devel - Oct 2007 - Reduce CPU Utilization make Network Perfromance better??

I''ve read "Performance Analysis and Opportunities for
Improvecment" in Xen Summit 2007.

J. Renato Santos et al. in HP Lab shows the siginificant CPU Cost of XenLinux.
(Compare to Linux)
I don''t understand the statistics about "CPU Utilization".
In my guess,
Xenoprof supported in Xen 3.0.3 is not available to activate Passive Domain Mode
profiling.
So if the benchmark server is hosted by domU, then the CPU utilization means the
summation of cycles
reported from dom0 and domU individually.
Comparing with the results hosted by Linux, it is three times longer.

I don''t know if the guess condition is wright.
But I still not figure out how the improvement benefit network throughput.
Is it because lower CPU utilization higher packet process rate??
Why not show network throught directly??


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

> I don''t know if the guess condition is wright.
> But I still not figure out how the improvement benefit network throughput.
> Is it because lower CPU utilization higher packet process rate??
Presumably because the CPU processing required to interact with the network 
can becomes a bottleneck; therefore improving the CPU efficiency should 
increase the throughput.  It''s been a long time since I read that
paper, but
when we did the first driver domains benchmarks in 2004 we were burning up a 
significant fraction of the CPU bandwidth of a 2-way Xeon in order to achieve 
line rate on gigabit ethernet.

Also, improving the CPU utilisation encountered in the course of network 
processing should make more CPU time available to other work, which would be 
good.
> Why not show network throught directly??
Presumably in the case of this paper the goal was to illustrate the kind of 
measurements that Xenoprof can make and the improvements in those 
measurements that can be achieved with targetted modifications to the 
codebase.

I agree it would have been interesting to see what effects on the network 
throughput were but I guess these would vary between different machines and 
network loads in any case.

Just my two cents ;-)

Cheers,
Mark

-- 
Dave: Just a question. What use is a unicyle with no seat?  And no pedals!
Mark: To answer a question with a question: What use is a skateboard?
Dave: Skateboards have wheels.
Mark: My wheel has a wheel!

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Just adding a few comments to Mark''s reply
>> -----Original Message-----
>> From: xen-devel-bounces@lists.xensource.com 
>> [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of 
>> Mark Williamson
>> Sent: Sunday, October 07, 2007 8:22 PM
>> To: xen-devel@lists.xensource.com
>> Cc: Brian Fang
>> Subject: Re: [Xen-devel] Reduce CPU Utilization make Network 
>> Perfromancebetter??
>> 
>> > I don''t know if the guess condition is wright.
>> > But I still not figure out how the improvement benefit 
>> network throughput.
>> > Is it because lower CPU utilization higher packet process rate??
>> 
>> Presumably because the CPU processing required to interact 
>> with the network can becomes a bottleneck; therefore 
>> improving the CPU efficiency should increase the throughput. 
>>  It''s been a long time since I read that paper, but when we 
>> did the first driver domains benchmarks in 2004 we were 
>> burning up a significant fraction of the CPU bandwidth of a 
>> 2-way Xeon in order to achieve line rate on gigabit ethernet.
>> 
  Yes, for gigabit NICs both Xen and linux can achieve
  line rate with modern machines
  As we move to 10Gb/s NICs the CPU will become the bottleneck,
  thus reducing the CPU cost should enable Xen to get better
  network throughput.
>> Also, improving the CPU utilisation encountered in the 
>> course of network processing should make more CPU time 
>> available to other work, which would be good.
>> 
  Right! Domains do more than just networking. Freeing cycles
  from the network stack should improve application performance.
  While the experiments are done using a micro network 
  benchamark (netperf), real applications will have to do real
  work in addition to network processing.
>> > Why not show network throught directly??
>> 
  Because it is not interesting. Both Xen and linux can
  saturate a 1 Gb/s NIC and thus achieve the same
  throughput.

>> Presumably in the case of this paper the goal was to 
>> illustrate the kind of measurements that Xenoprof can make 
>> and the improvements in those measurements that can be 
>> achieved with targetted modifications to the codebase.
>> 
  Actually the goal is to improve efficiency of Xen networking,
  to free cycles for the application and to scale to higher 
  network bandwidth (i.e. 10 Gb/s)
 
>> I agree it would have been interesting to see what effects 
>> on the network throughput were but I guess these would vary 
>> between different machines and network loads in any case.
>> 
>> Just my two cents ;-)
>> 
>> Cheers,
>> Mark
>> 
>> --
>> Dave: Just a question. What use is a unicyle with no seat?  
>> And no pedals!
>> Mark: To answer a question with a question: What use is a skateboard?
>> Dave: Skateboards have wheels.
>> Mark: My wheel has a wheel!
>> 
>> _______________________________________________
>> Xen-devel mailing list
>> Xen-devel@lists.xensource.com
>> http://lists.xensource.com/xen-devel
>> 
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Thanks for reply.

It''s really helpful to realize the reason about reducing CPU
utilization.
And it is told me that, Xen and Native Linux achive line rate under 1 Gps is 
truth.
Same as experimental I''ve done shows.
I perhaps need to obtain bandwith within 10 Gbps to make deference between 
Xen and Native Linux.

Sincerely,
Brian Fang
----- Original Message ----- 
From: "Santos, Jose Renato G" <joserenato.santos@hp.com>
To: "Mark Williamson" <mark.williamson@cl.cam.ac.uk>; 
<xen-devel@lists.xensource.com>
Cc: "Brian Fang" <m9429015@stmail.cgu.edu.tw>
Sent: Tuesday, October 09, 2007 1:57 AM
Subject: RE: [Xen-devel] Reduce CPU Utilization make Network 
Perfromancebetter??



Just adding a few comments to Mark''s reply
>> -----Original Message-----
>> From: xen-devel-bounces@lists.xensource.com
>> [mailto:xen-devel-bounces@lists.xensource.com] On Behalf Of
>> Mark Williamson
>> Sent: Sunday, October 07, 2007 8:22 PM
>> To: xen-devel@lists.xensource.com
>> Cc: Brian Fang
>> Subject: Re: [Xen-devel] Reduce CPU Utilization make Network
>> Perfromancebetter??
>>
>> > I don''t know if the guess condition is wright.
>> > But I still not figure out how the improvement benefit
>> network throughput.
>> > Is it because lower CPU utilization higher packet process rate??
>>
>> Presumably because the CPU processing required to interact
>> with the network can becomes a bottleneck; therefore
>> improving the CPU efficiency should increase the throughput.
>>  It''s been a long time since I read that paper, but when we
>> did the first driver domains benchmarks in 2004 we were
>> burning up a significant fraction of the CPU bandwidth of a
>> 2-way Xeon in order to achieve line rate on gigabit ethernet.
>>
  Yes, for gigabit NICs both Xen and linux can achieve
  line rate with modern machines
  As we move to 10Gb/s NICs the CPU will become the bottleneck,
  thus reducing the CPU cost should enable Xen to get better
  network throughput.
>> Also, improving the CPU utilisation encountered in the
>> course of network processing should make more CPU time
>> available to other work, which would be good.
>>
  Right! Domains do more than just networking. Freeing cycles
  from the network stack should improve application performance.
  While the experiments are done using a micro network
  benchamark (netperf), real applications will have to do real
  work in addition to network processing.
>> > Why not show network throught directly??
>>
  Because it is not interesting. Both Xen and linux can
  saturate a 1 Gb/s NIC and thus achieve the same
  throughput.

>> Presumably in the case of this paper the goal was to
>> illustrate the kind of measurements that Xenoprof can make
>> and the improvements in those measurements that can be
>> achieved with targetted modifications to the codebase.
>>
  Actually the goal is to improve efficiency of Xen networking,
  to free cycles for the application and to scale to higher
  network bandwidth (i.e. 10 Gb/s)
>> I agree it would have been interesting to see what effects
>> on the network throughput were but I guess these would vary
>> between different machines and network loads in any case.
>>
>> Just my two cents ;-)
>>
>> Cheers,
>> Mark
>>
>> --
>> Dave: Just a question. What use is a unicyle with no seat?
>> And no pedals!
>> Mark: To answer a question with a question: What use is a skateboard?
>> Dave: Skateboards have wheels.
>> Mark: My wheel has a wheel!
>>
>> _______________________________________________
>> Xen-devel mailing list
>> Xen-devel@lists.xensource.com
>> http://lists.xensource.com/xen-devel
>> 

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel