[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

答复: 答复: 答复: [Xen-devel] Credit scheduler vs SEDF scheduler


  • To: "'gaurav somani'" <onlineengineer@xxxxxxxxx>, <xen-devel@xxxxxxxxxxxxxxxxxxx>
  • From: "Zhiyuan Shao" <zyshao.maillist@xxxxxxxxx>
  • Date: Mon, 11 May 2009 17:54:28 +0800
  • Cc:
  • Delivery-date: Mon, 11 May 2009 02:55:21 -0700
  • Domainkey-signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=message-id:from:to:subject:date:mime-version:content-type:x-mailer :in-reply-to:thread-index:x-mimeole; b=x2uV9M3Ao/+3PA4oD8D41FNJ6TOe09M3GmB7jvY8feVQXU0Div6OhpEaTow86+qgJq qOW9xJamsbSZSAwHeHkud6oK+dh5SsuJ3MaWJdIOaqcSckZYIkwW8wWHgabbbg9bRYhc msYy98FUQw0+iIRYXrZ1kuUw+gyza2Ey06848=
  • List-id: Xen developer discussion <xen-devel.lists.xensource.com>
  • Thread-index: AcnRbaXLWTGIre+qQkmTMfKKm/fgTwAqpaUg

Some points I think that may help to your doubts:

1) how the scheduler of guest OS and VMM cooperate together.

Frankly, I also think this is a rather interesting topic to research, and currently working on that. However, before putting hands on this topic, we need some answers on several other tightly related problems. First problem is time keeping mechanism of Xen VMM. Not yet reading all the related codes of Xen, I guess for PV guests (do not consider the HVM case, since it is more complicated than PV), each time that a VCPU is scheduled in, one of the members of VCPU structure that stores current time will be updated (see struct vcpu_time_info), while the time precedes when the VCPU is actually is scheduled-out is simply “lost”. This will definitely cause problems by confusing the scheduler of the guest OS, however, hardly find documents that elaborate this problem or analyze what will happen in deep.

The second problem is load-balancing. Unfortunately, this is a more complicated problem. The guest OS (for the SMP guest cases) has its own load balancing strategy, and I think the bottom line is that it will not waste processing time. Taking an example here, running a single thread process (such as “dd if=/dev/zero of=./test.file bs=64 count=3200K”, it writes a 200MB file to virtual disk) inside a two-vcpu PV guest. The process originally running on top of single VCPU, but after a while, the VCPU is scheduled out (blocked and waiting for the data actually write to the virtual disk file), and another VCPU is scheduled in, the process will migrate to the other VCPU naturally. This case will result in extra overhead, and I got performance data to prove this. Besides understanding the load balancing strategy of guest OS, load balancing for scheduler of the VMM deserves also careful consideration and design. Current default scheduler of Credit actually do not doing very well on this point. The PCPU will steal VCPU from its neighbors once it found it is idle. Smarter solutions are needed in later schedulers (as George had mentioned).

2) what types of events are ocurring during scheduling

I think it is good for you to check the xentrace data for this problem.

3) xenmon problems.

I had not used that and can not give you any suggestions, sorry for that.

4) Xentop

Do not know what you mean for N/W?

 

Best,

Zhiyuan

 

 

 


发件人: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx] 代表 gaurav somani
发送时间: 2009年5月10 20:48
收件人: xen-devel@xxxxxxxxxxxxxxxxxxx
主题: Re: 答复: 答复: [Xen-devel] Credit scheduler vs SEDF scheduler

 

Hi list,

Some doubts.
(1) How the guest kernel scheduling strategies (linux kernel scheduling like DEADLINE and IO schedulers in guests) affect the VM Scheduling?

(2) I want to know what types of events are ocurring during scheduling- which domain is scheduled and what types of tasks it is running.

I am using xenmon-   some probs.

(1) it does not show me --iocount option correctly. it always shows iocount= 0, although i am running dd utility in one of the guset domain.
(2) when using xentop - it does not show any N/W data for domain 0.

am i missing something in configuration.


2009/5/7 Zhiyuan Shao <zyshao.maillist@xxxxxxxxx>

Believe your can use BCredit for this job. However, you can only set the attributes of a domain statically, which means the scheduler can not adjust its behavior according to the types of the work loads.

 

Remember there is some publication in proceedings of VEE 2009, which had addressed your problem. The solution continuously guests the performance characteristics of the application, and adjust the preemption decisions of the scheduler itself. You can take a look at that, but unfortunately, they did not provide the patch behind the solution. Besides, I am not sure that whether such solution is really helpful to solve real problems on real hardware, since it is based its decisions on possibilities.

 

Zhiyuan

 


发件人: gaurav somani [mailto:onlineengineer@xxxxxxxxx]
发送时间: 2009年5月6 17:09
收件人: Zhiyuan Shao
抄送: xen-devel@xxxxxxxxxxxxxxxxxxx
主题: Re: 答复: [Xen-devel] Credit scheduler vs SEDF scheduler

 

Hi Zhiyuan and George,

Thanks for the comments.

you are right i was getting spikes with more than 100ms and some times 400ms.

I will be trying with the patch u suggetsed.

one more aspect, can we classify schedulers according to applications they run.?

Like compute intensive on Credit and Latency intensive on sEDF or some trade off between them (In a cluster of servers).


Thanks

Gaurav

2009/5/6 Zhiyuan Shao <zyshao.maillist@xxxxxxxxx>

Hi,

 

Your experiment results are actually “expected”. There are several reasons make the results happen.

 

Firstly, Credit has long time slice (i.e., 30 millisec by default), and this make the observed close-to 30ms RTT to happen, actually, during such long RTT (the spike), the domain in test (domain 1 of your experiment) is just scheduled-out (actually, I guess the spikes illustrated in your figure 1 should read higher than 30ms!). At the same time, the scheduled-in vcpus (may executing the TEST program, which is cpu-hogs), and they will not give up their 30-ms time slice if they are not preempted.

 

Secondly, as regulated by Credit scheduler, if a VCPU has no more credits, it will not be BOOSTed even it do have some I/O requests to handle. This explains why your figure1, which has little spikes in the first 150 second, and crowed spikes follow. Since vcpu of domain 1 simply used out their credits, and can not preempt the other running vcpus of other domains, even they do have I/O operations to accomplish (i.e., PING).

 

Lastly, compared with Credit, SEDF generally has short time slices, and it has bigger chances to preempt current running vcpu when another vcpu found it has some I/O operations to be accomplished. (actually, I donot understand SEDF very well, but conclude these features by reading the trace data)

 

Actually, you can use BCredit patch to eliminate the spikes, I can guarantee that it will work very well! J And I agree with George that you should not use PING to justify whether a scheduler is good or not.

 

Best,

Zhiyuan

 

 


发件人: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx] 代表 gaurav somani
发送时间: 2009年5月5 17:26
收件人: xen-devel@xxxxxxxxxxxxxxxxxxx
主题: [Xen-devel] Credit scheduler vs SEDF scheduler

 

Hi list,

I am evaluating the scheduler behavior in xen.

I am using Xen 3.3.0
Dom0 and Dom1,2,3 and 4 all are opensuse 11.
I have one CPU Intensive TEST which has no. of arithmatic instruction in an infinite while() loop.
i am pinging domain1 with an external machine. and noting the RTT values.

i have below experiments
time (s)          domain state
0                 dom0,1,2,3,4 all idle
50               dom2 TEST started
100             dom3 TEST started
150             dom4 TEST started
200             dom0 TEST started
250             dom2 TEST stopped
300             dom3 TEST stopped
350             dom4 TEST stopped
400             dom0 TEST stopped

For these 400 seconds time, i have performed experiments with Credit and SEDF sceduler.
the configuration is


Credit configuration - weight 256, cap 0
Domain                  VCPU
0                                  2
1                                  2
2                                  2
3                                  2
4                                  2
all vcpu0s are pinned to pcpu0 and vcpu1s to pcpu1.


SEDF configuration - Period 10ms slice - 1.9ms
Domain                  VCPU
0                                  2
1                                  2
2                                  2
3                                  2
4                                  2
all vcpu0s are pinned to pcpu0 and vcpu1s to pcpu1.

the results of RTT values are attached herewith. the performance of Credit is very bad in comparison to SEDF in this scenario.
Please provide me some thought on it.


Thanks and Regards

Gaurav somani
M.Tech (ICT)
Dhribhai Ambani Institute of ICT,
INDIA

http://dcomp.daiict.ac.in/~gaurav






--
Gaurav somani
M.Tech (ICT)
Dhribhai Ambani Institute of ICT,
Gandhinagar
Gujarat, INDIA

onlineengineer@xxxxxxxx
gaurav_somani@xxxxxxxxxxxx
http://dcomp.daiict.ac.in/~gaurav
http://www.linkedin.com/in/gauravsomani




--
Gaurav somani
M.Tech (ICT)
Dhribhai Ambani Institute of ICT,
Gandhinagar
Gujarat, INDIA

onlineengineer@xxxxxxxx
gaurav_somani@xxxxxxxxxxxx
http://dcomp.daiict.ac.in/~gaurav
http://www.linkedin.com/in/gauravsomani

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

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.