Re: Performance test result between per-vhost kthread disable and enable

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Nov 23, 2010 at 10:13:43AM +0800, lidong chen wrote:
> I test the performance between per-vhost kthread disable and enable.
> 
> Test method:
> Send the same traffic load between per-vhost kthread disable and
> enable, and compare the cpu rate of host os.
> I run five vm on kvm, each of them have five nic.
> the vhost version which per-vhost kthread disable we used is rhel6
> beta 2(2.6.32.60).
> the vhost version which per-vhost kthread enable we used is rhel6 (2.6.32-71).

At this point, I'd suggest testing vhost-net on the upstream kernel,
not on rhel kernels. The change that introduced per-device threads is:
c23f3445e68e1db0e74099f264bc5ff5d55ebdeb

> Test result:
> with per-vhost kthread disable, the cpu rate of host os is 110%.
> with per-vhost kthread enable, the cpu rate of host os is 130%.

Is CONFIG_SCHED_DEBUG set? We are stressing the scheduler a lot with
vhost-net.

> In 2.6.32.60,the whole system only have a kthread.
> [root@rhel6-kvm1 ~]# ps -ef | grep vhost
> root       973     2  0 Nov22 ?        00:00:00 [vhost]
> 
> In 2.6.32.71,the whole system have 25 kthread.
> [root@kvm-4slot ~]# ps -ef | grep vhost-
> root     12896     2  0 10:26 ?        00:00:00 [vhost-12842]
> root     12897     2  0 10:26 ?        00:00:00 [vhost-12842]
> root     12898     2  0 10:26 ?        00:00:00 [vhost-12842]
> root     12899     2  0 10:26 ?        00:00:00 [vhost-12842]
> root     12900     2  0 10:26 ?        00:00:00 [vhost-12842]
> 
> root     13022     2  0 10:26 ?        00:00:00 [vhost-12981]
> root     13023     2  0 10:26 ?        00:00:00 [vhost-12981]
> root     13024     2  0 10:26 ?        00:00:00 [vhost-12981]
> root     13025     2  0 10:26 ?        00:00:00 [vhost-12981]
> root     13026     2  0 10:26 ?        00:00:00 [vhost-12981]
> 
> root     13146     2  0 10:26 ?        00:00:00 [vhost-13088]
> root     13147     2  0 10:26 ?        00:00:00 [vhost-13088]
> root     13148     2  0 10:26 ?        00:00:00 [vhost-13088]
> root     13149     2  0 10:26 ?        00:00:00 [vhost-13088]
> root     13150     2  0 10:26 ?        00:00:00 [vhost-13088]
> ...
> 
> Code difference:
> In 2.6.32.60,in function vhost_init, create the kthread for vhost.
> vhost_workqueue = create_singlethread_workqueue("vhost");
> 
> In 2.6.32.71,in function vhost_dev_set_owner, create the kthread for
> each nic interface.
> dev->wq = create_singlethread_workqueue(vhost_name);
> 
> Conclusion:
> with per-vhost kthread enable, the system can more throughput.
> but deal the same traffic load with per-vhost kthread enable, it waste
> more cpu resource.
> 
> In my application scene, the cpu resource is more important, and one
> kthread for deal with traffic load is enough.
> 
> So i think we should add a param to control this.
> for the CPU-bound system, this param disable per-vhost kthread.
> for the I/O-bound system, this param enable per-vhost kthread.
> the default value of this param is enable.
> 
> If my opinion is right, i will give a patch for this.

Let's try to figure out what the issue is, first.

-- 
MST