Fermi LAT Gamma-ray pulsar binary search "FGRPB1" - new app versions

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4265
Credit: 244921331
RAC: 16928
Topic 201332

We recently issued a couple of new app versions for FGRPB1:

- 1.01 is identical to the initial 1.01, but has the plan class "FGRPOLD" associated. This ensures that only "old" tasks are ran with that application

- 1.03 incorporates a fix for the final "follow-up" computation. The run-time of this application will be slightly longer (~2%). The result is incompatible (won't "validate") with that of 1.01, so we made sure that only tasks generated after this app version was published are ran with it.

- 1.04 fixes a problem with the Linux 64Bit build of 1.03

- 1.05 (out now!) saves some time with a pre-calculated "lookup table". Takes ~100MB more memory, but should speed up tasks by 15-25% (depends on the host)

BM

Jasper
Jasper
Joined: 14 Feb 12
Posts: 63
Credit: 4032891
RAC: 0

My first two 1.05 WUs

My first two 1.05 WUs finished last night. It took them about an hour less, out of ten, but they are not validated yet - I´m recovering (slowly) from a single invalid 12 days ago. It seems there is also quite some savings in additional run time, but I´d need to see more over longer time, as that depends on what I´m doing myself with it when not sleeping.

It is not as much less as announced for me, but much appreciated nevertheless. A whole hour out of 10 makes it quite significant in my book. 

AgentB
AgentB
Joined: 17 Mar 12
Posts: 915
Credit: 513211304
RAC: 0

Bernd Machenschalk wrote:-

Bernd Machenschalk wrote:

- 1.05 (out now!) saves some time with a pre-calculated "lookup table". Takes ~100MB more memory, but should speed up tasks by 15-25% (depends on the host)

Thanks Bernd, I don't have a significant sample to make good comment yet but the first 1.05 task was about 40% slower on this host.   I have others in progress which i think may be faster (not on that host)

FGRPB1 on host 1905468

my partner on that task has two tasks which look about 5% slower.

FGRPB1 on host 9891958

And his partner a E3-1220 v3 Xeon also looks to be about 40-50 % slower.

FGRPB1 on host 12245654

I've tried to sort the links so that the 1.05 float to the top.

 

AgentB
AgentB
Joined: 17 Mar 12
Posts: 915
Credit: 513211304
RAC: 0

On this Xeon  E5-2660 host

On this Xeon  E5-2660 host 12242223

Application     N     Average      Median     Mininum      Maximum   StdDeviation
FGRPB1v1.00  6427     37969.5      38040.67   16175.79     41602.40  558.528
FGRPOLD        68     37920.9      38093      36846.11     38321.14  394.326
FGRPSSE        14     42919.7      42947.6    42673.05     43098.47  127.486
FGRPSSEv1.04  102     38768        38870.8    37339.38     39341.85  352.08
FGRPSSEv1.05   34     39422.7      39427.4    39258.41     39707.46   98.146

 On this i5-4690k host 11905468.

Application     N     Average      Median      Mininum       Maximum StdDeviation
FGRPB1v1.00   395     10610.8     10456.41      671.54      14646.30 1059.96
FGRPOLD        12     10464.8     10460.7     10375.99      10568.47   48.1537
FGRPSSE        21     15897.7     16574.71     5057.27      16725.81 2487.38
FGRPSSEv1.04    6     10847.3     10865.1     10686.62      11010.61  111.035
FGRPSSEv1.05    5     15653.5     15622.37    15559.85      15893.64  123.521

 

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4265
Credit: 244921331
RAC: 16928

I declared the 1.05 version

I declared the 1.05 version as "Beta", so you can decide yourself whether to run it or not. So far I really don't know on which host it's faster or slower.

BM

Jasper
Jasper
Joined: 14 Feb 12
Posts: 63
Credit: 4032891
RAC: 0

That does not seem to work

That does not seem to work for me, as I´d prefer 1.05 of course here. The selection seems OK but preference is given to 1.03, much longer for me:

2016-09-26 12:47:22.0980 [PID=23112]    [version] Checking plan class 'FGRPOLD'
2016-09-26 12:47:22.0980 [PID=23112]    [version] WU#256278444 too new
2016-09-26 12:47:22.0980 [PID=23112]    [version] Checking plan class 'FGRPSSE'
2016-09-26 12:47:22.0980 [PID=23112]    [version] plan class ok
2016-09-26 12:47:22.0980 [PID=23112]    [version] Checking plan class 'FGRPSSE-Beta'
2016-09-26 12:47:22.0980 [PID=23112]    [version] plan class ok
2016-09-26 12:47:22.0980 [PID=23112]    [version] Best version of app hsgamma_FGRPB1 is 1.03 ID 839 FGRPSSE (3.21 GFLOPS)

I am running 7.6.22 on OS X. 

 
Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4265
Credit: 244921331
RAC: 16928

Sorry. Should work now.

Sorry. Should work now.

BM

pustota
pustota
Joined: 31 Aug 16
Posts: 1
Credit: 7837601
RAC: 0

Hi! I started getting errors

Hi!

I started getting errors as a result of the calculation tasks in version 1.05

Also i had Boinc crash couple times with this tasks. i didn't get any errors on prev. task versions (1.01,1.03)

How can i fix this?  my computer is 12418567

Thanks.

 

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109379382859
RAC: 35973647

Bernd Machenschalk wrote:I

Bernd Machenschalk wrote:

I declared the 1.05 version as "Beta", so you can decide yourself whether to run it or not. So far I really don't know on which host it's faster or slower.

If a volunteer has allowed beta apps for the purposes of using a particular GPU app, then there wont be an opportunity to use the non-beta CPU app, will there???  If a particular host is using a beta GPU app, wont it also be using the beta CPU app???

In my case, I have allowed beta apps but would prefer to use the non-beta version of the FGRPB1 app.  As a quick example of why, on a Haswell based machine (G3258 CPU + HD7850 GPU) running 64 bit Linux, the 1.04 app was taking around 14.5 ksecs/task whilst the 1.05 app now takes >20 ksecs.  This isn't an isolated example.  I think a lot (if not most) of my hosts are in the same boat.

I don't have time to do a proper analysis and I find it much harder to gather the needed information on the new website.  Without the ability to select results for a particular science run, it's just too tedious and frustrating.

 

Cheers,
Gary.

Jasper
Jasper
Joined: 14 Feb 12
Posts: 63
Credit: 4032891
RAC: 0

I´m not sure, maybe it was

Bernd Machenschalk wrote:

Sorry. Should work now.

I´m not sure, maybe it was just my bad. I enabled Beta testing yesterday morning, saved and that was it, thinking it would take precedence over anything else. After a while, I realised this was perhaps not true: I was still running under local preferences, where Beta testing is not an option. So I activated web preference instead, yesterday evening...

Anyway, I got two 1.05 WUs meanwhile, around midnight. Should finish shortly. Embarassed 

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4265
Credit: 244921331
RAC: 16928

Gary Roberts wrote:Bernd

Gary Roberts wrote:

Bernd Machenschalk wrote:

I declared the 1.05 version as "Beta", so you can decide yourself whether to run it or not. So far I really don't know on which host it's faster or slower.

If a volunteer has allowed beta apps for the purposes of using a particular GPU app, then there wont be an opportunity to use the non-beta CPU app, will there???  If a particular host is using a beta GPU app, wont it also be using the beta CPU app???

Sorry. You might agree that it's pretty hard to please everyone.

It would help, I guess, to find out on which systems the 1.05 runs faster and on which it's slower, so I can implement an automatic selection independently of the manual "Beta" one.

BM

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.