| 1 |
Sorry.. I should have clarified that.. That is not our 'official, |
| 2 |
complete' results.. just data that was thrown on as benchmarks |
| 3 |
completed. |
| 4 |
|
| 5 |
The bars represent percentage comparison to the highest scoring |
| 6 |
result. Aka: #1 is always 100%, the next is showing a percentage of |
| 7 |
that. The reason we did it that way was because some of the results |
| 8 |
where in the 76-78 range and some where in the several hundred |
| 9 |
thousand range.. We couldn't make a decent chart of all the results |
| 10 |
without normalizing them all to a percentage rate. |
| 11 |
|
| 12 |
Anyways.. the raw numbers represent an index from a base system that |
| 13 |
was setup when UnixBench was written. A score of 1 = just as fast as |
| 14 |
that system. 2 = some linear value faster.. etc. That system was a |
| 15 |
386 if I remember correctly so most scores are much much higher than |
| 16 |
1 any more. Things like memory performance and stuff have come light |
| 17 |
years. |
| 18 |
|
| 19 |
So the short of it is that higher is always better, and this is only |
| 20 |
a 'best as I personally' could do comparison. |
| 21 |
|
| 22 |
After our conversion things went together much smoother and now |
| 23 |
maintenance is fairly painless. I spend all of my time setting up our |
| 24 |
Apple cluster (The Gentoo ppc64 performance just wasn't good enough |
| 25 |
to wow the management into using it =) |
| 26 |
|
| 27 |
One big advantage of Gentoo is the ease in which new programs/ |
| 28 |
libraries can be installed. I have written dozens of ebuilds for all |
| 29 |
the programs we use here in order to simplify installation and |
| 30 |
administration. |
| 31 |
|
| 32 |
Now to install a new program I just have to build it on a node using: |
| 33 |
echo "emerge -b <program>" | qsub |
| 34 |
|
| 35 |
Then I install it on everything else using: |
| 36 |
emerge -K <program> ; pdsh -a emerge -K <program> |
| 37 |
|
| 38 |
|
| 39 |
(Though I am so lazy that I even scripted the whole thing so emerging |
| 40 |
a program on my cluster involves: |
| 41 |
mass_emerge <program> |
| 42 |
|
| 43 |
=) |
| 44 |
|
| 45 |
|
| 46 |
|
| 47 |
On Apr 11, 2006, at 2:07 PM, Donnie Berkholz wrote: |
| 48 |
|
| 49 |
> Brady Catherman wrote: |
| 50 |
>> I just converted out 134 node cluster from RedHat Enterprise Linux |
| 51 |
>> v4 to |
| 52 |
>> Gentoo 2005.1 |
| 53 |
>> |
| 54 |
>> Before and after the conversion we ran UnixBench as part of the |
| 55 |
>> Beowulf |
| 56 |
>> Performance Suite. |
| 57 |
>> |
| 58 |
>> There is a graphical representation of the results here: |
| 59 |
>> http://oceanus.ibest.uidaho.edu/~bcatherm/benchmark.png |
| 60 |
>> |
| 61 |
>> That shows Gentoo and Redhat on x86 hardware, and Gentoo and Macos |
| 62 |
>> on an |
| 63 |
>> dual XServe G5. |
| 64 |
> |
| 65 |
> The graph is a bit confusing because you essentially need to |
| 66 |
> compare the |
| 67 |
> two inner bars, and the two outer bars. It might be clearer with the 2 |
| 68 |
> x86 setups adjacent, and the 2 ppc setups adjacent. It's also a bit |
| 69 |
> unclear to me whether high values are universally good, or low values, |
| 70 |
> or whether it varies from test to test. From what I could find, it |
| 71 |
> looked like high was good for file copy and low for the rest, but |
| 72 |
> maybe |
| 73 |
> they've been modified so high is always good. |
| 74 |
> |
| 75 |
> Could you provide some more detailed hardware specs and `emerge info`? |
| 76 |
> I'd really like to put this data up somewhere on |
| 77 |
> gentoo.org/proj/en/cluster/ -- with your name etc withheld if |
| 78 |
> necessary. |
| 79 |
> |
| 80 |
> It would be interesting to do some performance analysis to see |
| 81 |
> where any |
| 82 |
> slowdowns come from and try to get things up to speed. |
| 83 |
> |
| 84 |
> Thanks, |
| 85 |
> Donnie |
| 86 |
> |
| 87 |
|
| 88 |
-- |
| 89 |
gentoo-cluster@g.o mailing list |
Archives
