I will be re-doing the whole
test with a 4GB hard limit on the memory (for both the native case and the HVM
case) and also with 2 VCPU allocated for Dom0.
I don’t know what the difference
is in our environments. Maybe he was looking at overall time spent? I am
interested in the maximum jobs/minute which can be an indication of how much
horsepower we can get out of a guest VM.
Any answers for my other
questions?
Best regards,
---Kayvan
From: Xu, Anthony
[mailto:anthony.xu@xxxxxxxxx]
Sent: Thursday, January 31, 2008 5:11 PM
To: Kayvan Sylvan; xen-ia64-devel
Subject: RE: [Xen-ia64-devel] HVM Multi-Processor Performance followup
You
can see the drop in performance starts to get really bad at about 9 CPUs and
beyond
If
you increase guest vCPU number, the bottleneck may be dom0 vCPU number( only
1vCPU for dom0).
You
can try configure two/four vCPU for dom0, the performance may be back.
Alex
said there are ~70% degradation on RE-AIM7,
Your
test result seems much better than his.
What's
the difference of your test environment?
From: xen-ia64-devel-bounces@xxxxxxxxxxxxxxxxxxx
[mailto:xen-ia64-devel-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of Kayvan
Sylvan
Sent: 2008年2月1日 8:57
To: xen-ia64-devel
Subject: [Xen-ia64-devel] HVM Multi-Processor Performance followup
Hi everyone,
A follow-up on the multiprocessor performance benchmark on
HVM guests.
We ran the RE-AIM7 benchmarks on a 5-cell (40 CPU) machine
and a single-cell 8-cpu NEC machine.
Here are the jobs per minute maximums.
You can see the drop in performance starts to get really bad
at about 9 CPUs and beyond.
Questions:
1.
What can I do to help improve this situation?
2.
Are there any other experiments I can run?
3.
What tools/profilers will help to gather more data here?
I am very interested in helping to solve this problem!
Thanks for your ideas and suggestions.
Best regards,
---Kayvan
|
|
|
|
|
|
|
Xen performance comparison on
5-Cell NEC machine (each cell with 4 dual-core Itaniums)
|
|
|
|
|
|
|
|
|
|
CPUs
|
Native
Jobs/Min
|
HVM
Jobs/Min
|
Overhead
|
|
|
|
|
1
|
2037
|
1791
|
12.08%
|
|
|
|
|
2
|
4076
|
3615
|
11.31%
|
|
|
|
|
3
|
6090
|
5221
|
14.27%
|
|
|
|
|
4
|
8118
|
6839
|
15.76%
|
|
|
|
|
5
|
10119
|
8404
|
16.95%
|
|
|
|
|
6
|
12037
|
9949
|
17.35%
|
|
|
|
|
7
|
14106
|
11095
|
21.35%
|
|
|
|
|
8
|
15953
|
12360
|
22.52%
|
|
|
|
|
9
|
18059
|
13201
|
26.90%
|
|
|
|
|
10
|
20170
|
13742
|
31.87%
|
|
|
|
|
11
|
21896
|
13694
|
37.46%
|
|
|
|
|
12
|
24079
|
13331
|
44.64%
|
|
|
|
|
13
|
25992
|
12374
|
52.39%
|
|
|
|
|
14
|
28072
|
11684
|
58.38%
|
|
|
|
|
15
|
29931
|
11032
|
63.14%
|
|
|
|
|
16
|
31696
|
10451
|
67.03%
|
|
|
|
|
|
|
|
|
|
|
|
|
The guest OS was CentOS-4.6
with 2GB of memory,
|
|
|
|
|
running under a Dom0 that was
limited to 1 VCPU.
|
|
|
|
|
|
|
|
|
|
|
Xen performance comparison on
1-Cell NEC machine (4 dual core Itanium Montecito)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
CPUs
|
Native
Jobs/Min
|
HVM
Jobs/Min
|
Overhead
|
|
|
|
|
|
|
|
1
|
2037
|
1779
|
12.67%
|
|
|
|
|
|
|
|
2
|
4067
|
3619
|
11.02%
|
|
|
|
|
|
|
|
3
|
6097
|
5344
|
12.35%
|
|
|
|
|
|
|
|
4
|
8112
|
7004
|
13.66%
|
|
|
|
|
|
|
|
5
|
10145
|
8663
|
14.61%
|
|
|
|
|
|
|
|
6
|
12023
|
10213
|
15.05%
|
|
|
|
|
|
|
|
7
|
14083
|
11249
|
20.12%
|
|
|
|
|
|
|
|
8
|
16182
|
12969
|
19.86%
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The guest OS was CentOS-4.6
with 2GB of memory,
|
|
|
|
|
|
|
|
running under a Dom0 that was
limited to 1 VCPU.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
