1 |
I just converted out 134 node cluster from RedHat Enterprise Linux v4 |
2 |
to Gentoo 2005.1 |
3 |
|
4 |
Before and after the conversion we ran UnixBench as part of the |
5 |
Beowulf Performance Suite. |
6 |
|
7 |
There is a graphical representation of the results here: |
8 |
http://oceanus.ibest.uidaho.edu/~bcatherm/benchmark.png |
9 |
|
10 |
That shows Gentoo and Redhat on x86 hardware, and Gentoo and Macos on |
11 |
an dual XServe G5. |
12 |
|
13 |
And here is the tables: |
14 |
Gentoo_x86% Gentoo_ppc6% MacOS % RedHat_x86% Gentoo_x86 Gentoo_ppc64 |
15 |
MacOS RedHat_x86 |
16 |
Dhrystone 2 using register variables 74.0980112 80.10593524 100 |
17 |
76.91090096 4248495.5 4592966.8 5733616.1 4409775.8 |
18 |
Double-Precision Whetstone 52.82766909 58.5163918 100 65.96781732 |
19 |
791.2 876.4 1497.7 988 |
20 |
Execl Throughput 100 25.90300029 16.30133994 0.680891349 2746.4 711.4 |
21 |
447.7 18.7 |
22 |
File Copy 1024 bufsize 2000 maxblocks 100 56.72337653 34.19748972 |
23 |
67.2042237 264034 149769 90293 177442 |
24 |
File Copy 256 bufsize 500 maxblocks 100 58.5222193 34.24796214 |
25 |
72.20220878 76060 44512 26049 54917 |
26 |
File Copy 4096 bufsize 8000 maxblocks 100 61.8005082 21.06610594 |
27 |
60.05659416 574971 355335 121124 345308 |
28 |
Pipe Throughput 100 83.1021079 36.09782829 88.49157048 541614.3 |
29 |
450092.9 195511 479283 |
30 |
Pipe-based Context Switching 100 39.30622691 68.55390149 na 123547.6 |
31 |
48561.9 84696.7 na |
32 |
Process Creation 100 40.99792243 9.929224748 72.82264808 11600.1 |
33 |
4755.8 1151.8 8447.5 |
34 |
Shell Scripts (8 concurrent) 100 49.33560892 26.2099232 9.216140436 |
35 |
820.3 404.7 215 75.6 |
36 |
System Call Overhead 87.00151942 100 28.1422103 79.61936462 481097 |
37 |
552975.4 155619.5 440275.5 |
38 |
|
39 |
|
40 |
|
41 |
|
42 |
|
43 |
|
44 |
On Apr 11, 2006, at 12:04 PM, Simon-Nicolas Roth wrote: |
45 |
|
46 |
> Hi, |
47 |
> |
48 |
> A small introduction: |
49 |
> We currently own 2 small clusters. One with 42 opteron diskless |
50 |
> nodes (adelie |
51 |
> gentoo linux) and another one with 24 pentium 4 diskless nodes |
52 |
> (based on |
53 |
> Kyron's wiki http://wiki.neuralbs.com/index.php/ |
54 |
> Gentoo_Diskless_Client). |
55 |
> |
56 |
> We have seen a nice upgrade in perfomance when we changed the 24 |
57 |
> nodes cluster |
58 |
> from Redhat to Gentoo. |
59 |
> |
60 |
> The question: |
61 |
> We are in discussion to make a cluster with around 1024 nodes for |
62 |
> parallel |
63 |
> scientific computing. I want to convince that Gentoo would be |
64 |
> better on this |
65 |
> cluster than RedHat! Most people around this project don't believe |
66 |
> in Gentoo |
67 |
> (obviously because they have never done anything with Gentoo). |
68 |
> |
69 |
> Does anyone have really solid arguments/articles/proofs about the |
70 |
> performance |
71 |
> of Gentoo on clusters to help me convice the RedHat guys? |
72 |
> |
73 |
> Thanks |
74 |
> Simon |
75 |
> -- |
76 |
> gentoo-cluster@g.o mailing list |
77 |
> |