How to compare WU's/PC's?
14 Sep 2009 9:37:54 UTC
Topic 194535
(moderation:
Hello,
Is there a way to see (compare) if WU’s are the same in size or time needed to complete. I think a fast PC does a job more quickly but that is not what I am experiencing.
I have i7 2,66GHz with over 7000m ops/sec and Q 2,40GHz with 5000m ops/sec. The latter uses around 16000 seconds to complete and the i7 around 25000s seconds for an S5 #5 job. This is strange to me, as I would think the i7 is faster.
Any ideas?
Greetings from
TJ
How to compare WU's/PC's?
)
Are they both hierarchical or astrobinarypulsar?
Tullio
Hi! The S5R5 work units are
)
Hi!
The S5R5 work units are indeed varying in runtime (quite significantly). To some degree, this is reflected by the credits claimed for the WUs, so that "longer" WUs will tend to claim more credits based on a runtime estimation done when the WU is created. This estimation is not perfect, but it's better than giving identical credit to long and short WUs.
CU
Bikeman
Perhaps you'll find some old
)
Perhaps you'll find some old threads interesting:
S5R5 Performance Analysis
Skygrid file analysis
Ready Reckoner Area (this one points at a message with interesting plots towards the bottom of the thread)
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
Thanks so far To Tullio:
)
Thanks so far
To Tullio: both hierarchical.
To Gundolf: I have a lot to reed.
To Bikeman: I see the difference in credit, that’s fair. Can I find that somewhere? When I compare my machines and some of the bosses (therefore hidden), it seems that the quad core is faster than the i7. Perhaps the i7 gets the “longer†WU’s and the Q the "shorter". That, I would like to “see†somewhere.
Greetings from
TJ
RE: Thanks so far To
)
It's simple, with http://einstein.phys.uwm.edu/hosts_user.php you see a table of the hosts under your account. Click on the "tasks" links to see the tasks for those hosts, and there is a claimed credits column for those tasks that are already reported. If you average over several tasks, you should get an approximate idea about the efficiency of the host in (credits / CPU second).
CU
Bikeman
I have an i7 920 too and some
)
I have an i7 920 too and some Quad Core Q9550 in use, indeed the Quad is
faster for a single task, but it computes only 4 tasks at once, due to
hyperthreading the i7 computes 8 with some losses due to the lower
frequency and the hyperthreading - over all my estimate is, that the i7
computes 40 to 70 percent more in the same time than the Quad Core, current
relation of average credits is for example 4652/2881.
This happened too with hyperthreading in Pentium 4 already, therefore this is
no surprise.
With i7 you can get one single task faster, if you use only one CPU and
nothing else on the computer - because i7 has this turbo feature ;o)
To avoid hyperthreading causes an acceleration too, but finally you will get
less tasks per month as with hyperthreading and using all 8 CPUs.
Maybe losses could be reduced, if it would be possible to address one task
exclusively to exactly one CPU. Normally the processes jump from
time to time from one CPU to the other including some losses. But this happens
on the Quad Cores too. But I don't know, how to change this.
To get a meaningful comparison, you just have to run both machines for about
a month without anything else on it and to compare the average credits.
RE: Maybe losses could be
)
Crunch3r released a boinc-client (6.1.0.32/64 V5 based on 5.10.X) with activated affinity and an app should be pined to one core but you can not pin to a specific core. I think the affinity is not needed anymore on newer native quads like Core i3/5/7 or all AMD because they do not have a fsb-interface. Each core can talk direct to another core without the connection over the slower fsb.
Sorry it’s me again, but I
)
Sorry it’s me again, but I still don’t understand.
I did already a check of the tasks and found around 17000 seconds (mostly less) for the Q6600 @2.40GHz and around 24000 seconds for the i7 @2.67GHz. Both for S5, #5 searches.
Both run 24/7 and the Q for more than a year and the i7 for 4 months. The i7 is also running the Einstein CUDA-app, but that is only using one core.
If I see this than I should think the Q6600 is more efficient than the i7 as it is approx. 10000 seconds faster for one WU. It can only run 4 and the i7 run 8 (currently only one as it got 10 CUDA tasks yesterday). Or the WU’s the Q6600 gets are “smallerâ€. That the Q6600 is about 10000 seconds “faster†confuses me. I would like to build another pc to run, but I don’t know what to make. I should think a quad core. Or do I misunderstand. Hope someone can advice. Thanks.
Greetings from
TJ
Well compared to my notebook
)
Well compared to my notebook Duo P7350, 2GHz (does CUDA beta too) even this is
faster per task as the i7, but including the GPU task is is only able to
do two or three tasks at once, therefore over all it computes less tasks per
month as the i7, average P7350+CUDA/i7: 1450/4800.
Due to my observation these benchmark results (mips, flops) derived from boinc
are only random fun and have not much to do with the over all computing speed
for one task or for tasks per month. Maybe this results from hyperthreading
and the turbo feature too, because boinc stops computation to do the benchmark
tests. If you are lucky, you will see a larger number if the turbo feature
is activated for the benchmark test.
What I can see are 19000s to 29000s per task on the i7 and with the Q9550
2.83GHz I see about 11000s to 25000s, for the P7350 2GHz I see about 18000s
to 24000s.
Therefore your 25000s and 16000s fit into this range and I think, there is
no interesting difference to what I observed.
As already mentioned, maybe you get it faster per task, if you switch off
hyperthreading (BIOS settings?), but I think, finally this will reduce the
tasks per month, even if the computing time per task might be shorter.
Not sure, if this works too, if you just limit boinc to 4 CPUs, you might still
have losses due to the possibility of hyperthreading, even if not used by boinc.
Because the range of computing times per task is pretty broad, you need a lot
of tasks to average and compare two computers - but why to do this?
The design of the i7 is different from the quad core or the notebook processor,
therefore all of them have some advantages or disadvantages. One can expect,
that for the i7 and maybe the notebook processor the 'tasks per kJoule' rate
is slightly better than that of the quad core and that the i7 finally computes
much more per month than the others due to hyperthreading and the turbo feature.
You can look into the BIOS to be sure, that this is switched on.
And if both of your computers already run for more than one month, you simply
can look in your profile on 'your computers' and compare the average credit,
if that of the i7 is much larger than that of the quad, it indeed computes more
per month but is slower per task.
RE: Hello, Is there a way
)
The i7 is indeed slower despite its slightly higher clock speed, if you have HT turned on. On an i7 with HT turned on you have 2 cores competing for the same cache memory. The Q's have dedicated cache for each core (and more of it).
I have upgraded from Q6600's to i7's and think of the i7 as equivilent to a 6 core machine (performance-wise).
BOINC blog