| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: Mapping Cancer Markers Forum
Thread: Haswell is faster |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 26
|
|
| Author |
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges: 
|
I am going to throw this out for what it is worth, rather than as a rigorous scientific comparison, but my preliminary tests are that Haswell does better on MCM by more than you would expect as compared to Ivy Bridge.
(Note that all speeds are stock, which are about the same for both CPUs, and running Win7 64-bit in all cases.) My main PC has four virtual cores of a Haswell i7-4771 running MCM, and has been finishing the work units in about 2 hours. My dedicated PC has six cores of an Ivy Bridge i7-3770 on MCM, and has been finishing the work units in about 2 hours 25 minutes on average. So Haswell appears to be about 20% faster, which is not earth-shaking but a little more than I expected. However, it is the outliers that I find curious. The longest ones on the Ivy Bridges take about 4 hours 30 minutes, whereas the longest on the Haswell are only about 3 hours 15 minutes, and there are fewer of them. Now the Haswell has been running only 3 days, so it might pick up longer ones later, but the Ivy Bridge was running only 1 1/2 days when it picked up its larger range. It might be that since I am running more cores on the Ivy Bridge, the cache has filled up, but it could also be that the more advanced instructions on Haswell (or some other architectural change) have been put to good use. I won't be using the Ivy Bridge on MCM any more, so it is somewhat academic to me at this point, but others might want to verify it and try to figure out what is going on. |
||
|
|
duanebong
Advanced Cruncher Singapore Joined: Apr 25, 2009 Post Count: 134 Status: Offline Project Badges: 
|
Both cpus have 4 cores (and 8 hyperthreads). Since you run 4 instances on the Haswell, they will usually have access to a full core each. On the Ivybridge you're running 6 instances - so 4 of them will be "sharing" a core at any one time. So to me it seems about right for the IvyBridge.
---------------------------------------- |
||
|
|
Jim1348
Veteran Cruncher USA Joined: Jul 13, 2009 Post Count: 1066 Status: Offline Project Badges:
|
The light load might be part of it, but there is a little more to the story. On both PCs, I am also running Folding@home on two AMD cards (two HD 7870s on the Ivy Bridge, and two HD 7790s on the Haswell). They each reserve a core for themselves, so that is really 6 virtual cores in use on the Haswell, but all 8 on the Ivy Bridge. On the other hand, Folding does not use much of each core that it reserves (less than 10% of each core). But that still leaves two cores free on the Haswell with nothing to do except desktop use (but that it intermittent), so that could account for it.
I suppose there is no point speculating without doing a more controlled test, which maybe someone else will do. Thanks for your input. |
||
|
|
ryan222h
Senior Cruncher Joined: Sep 4, 2006 Post Count: 425 Status: Offline |
There's too many variables to make a determination with any accuracy, Starting with the work units all being different sizes. I have sandy bridge and haswell systems and can only say that haswell is somewhat faster.
----------------------------------------Perhaps doing a BOINC benchmark on both machines, with nothing else running would give a good idea. If there's a simple way to run the exact same work unit (like make a copy) on multiple systems I would like to know how.  |
||
|
|
Speedy51
Veteran Cruncher New Zealand Joined: Nov 4, 2005 Post Count: 1326 Status: Recently Active Project Badges: 
|
Jim thanks for shearing your findings. Out of interest have you found any increase in one times. As of 8 January overall run-time was around 5.59 hours per task. Congrats on your gold badge. I need about 9 hours under 10 days run-time to get to my gold badge. I should complete this in a couple of days as I have been averaging 5 days per calendar day
---------------------------------------- |
||
|
|
Barnsley_Tatts
Senior Cruncher Joined: Nov 3, 2005 Post Count: 291 Status: Offline Project Badges:
|
You could always run the WCG benchmarks for each PC to compare them!
---------------------------------------- |
||
|
|
Sgt.Joe
Ace Cruncher USA Joined: Jul 4, 2006 Post Count: 7846 Status: Offline Project Badges:
|
Another way to compare them would be to track the results of about 10 days work on each system of a particular project. This would give a big enough base to compare each system, even though individual WU would vary. Using a calculation such as average points per hour over that time span should give you an indication of how the systems compare.
----------------------------------------Cheers
Sgt. Joe
*Minnesota Crunchers* |
||
|
|
yojimbo197
Advanced Cruncher Joined: Jun 30, 2012 Post Count: 83 Status: Offline Project Badges:
|
This reminds me of something I wish WCG would add. I wish they'd have a store of well coded WU's from each project that people could run on their systems to benchmark.
---------------------------------------- |
||
|
|
Sgt.Joe
Ace Cruncher USA Joined: Jul 4, 2006 Post Count: 7846 Status: Offline Project Badges:
|
This reminds me of something I wish WCG would add. I wish they'd have a store of well coded WU's from each project that people could run on their systems to benchmark. That would be nice, but I use WCGDAWS, which gives me some nice averages across machines and by project if I want to do it that way. Cheers
Sgt. Joe
*Minnesota Crunchers* |
||
|
|
yojimbo197
Advanced Cruncher Joined: Jun 30, 2012 Post Count: 83 Status: Offline Project Badges:
|
This reminds me of something I wish WCG would add. I wish they'd have a store of well coded WU's from each project that people could run on their systems to benchmark. That would be nice, but I use WCGDAWS, which gives me some nice averages across machines and by project if I want to do it that way. Cheers Right, but different people/rigs doing the same tasks would allow for a more accurate comparison and would not depend on batch variations in WU's for a particular project. [Edit 1 times, last edit by yojimbo197 at Jan 9, 2014 9:58:44 PM] |
||
|
|
|