5. BIONC Pentathlon 2014

[Rattledagger]

As it is said on the front page:

I have run the project for some time, and had no trouble at all.

Yes... but, how about some info about the wu's, like how long are they (5 minutes or 5 hours or 5 days)? Is it a limit on #wu's in progress on computer or per core? Do all wu's need wingman to validate, or do all wu's initially need validation but afterwards doesn't need wingman?

How long until deadline on the wu's? How much memory & disk-space does each wu need? Upload/download-usage? How is checkpointing-frequency?

Is there any special considerations like with WCG's Clean Energy Project 2, where if you runs "too many" at once the hd can't keep-up and you get "no heartbeat" possibly losing many hours of crunching in the process?

Also, is there any large speed-differences between platforms, example is Amd-cpu's 3x faster than intel or is Amd total crap in this project? Is Linux 2x faster than windows on same computer or are windows 2x faster? Does the 64-bit applications have a significant advantage over 32-bit, or is possibly the 32-bit significantly faster?


Now I don't expect answers to everything, but having an idea about deadline is highly relevant when it comes to pre-caching of work (or not if the deadline is only 2 days), a rough idea about wu-size in case the initial estimates is totally wrong so atleast doesn't cache much too few wu's, and disk-usage in case needs to make more space available (and guard against most of the work erroring-out due to hd being full).

 

[GLeeM]

avg runtime of last 100 results - 2.5 hours
- my 4770@4.2 with HT takes ~5000 seconds
- my i7 920@3948 with HT takes ~8000 seconds
looks like WUs need one wingman
deadline is about one week
memory is less than 30K
I have to be careful with upload/download because of satellite ISP and I don't remember these using much EDIT: 5KB download

hope this helps

 

[Ken g6]

Ooh, POGS is my project of choice for my ARM tablet! It takes ~12 hours per WU, and it gives better credit than any other Android project I've found. It also takes some memory, so I don't run it on my phone - but mainly because my phone is extremely cheap.

 

[Rattledagger]

Thanks for the info, GLeeM.

After finally having time to add the project and filling-up the cache, it seems a large 2-file download totalling around 18 MB is used for many wu's, so atleast in my experience roughly 700 wu's equals roughly 250 MB disk-usage (including application).

 

[Assimilator1]

Hmm, I totally forgot about this! :$, did anyone sign the team in?

[edit] ah yes, at least Petrus is on the ball

Skynet looks interesting btw , how far in advance do we need to switch?

 

[Rattledagger]

Skynet looks interesting btw , how far in advance do we need to switch?

You should preferably queue-up work ASAP, since with a 7-day deadline and race-start on the 5th. anything you can crunch now and report in 5 days will be an advantage.

 

[Assimilator1]

Wouldn't only credit turned in from the 5th count though?

 

[GLeeM]

Wouldn't only credit turned in from the 5th count though?

It is not like PrimeGrid races. You can crunch early but turn them in after the start and all will count.

If you get a bunch of WUs now, turn off network communication and then after the start send them in you will be ahead.

Trouble is: how do you get enough GPU WUs also? Do people use two clients to do this? One for pogs with communication turned off and another for GPU with communication turned on?

 

[Rattledagger]

Trouble is: how do you get enough GPU WUs also? Do people use two clients to do this? One for pogs with communication turned off and another for GPU with communication turned on?

Where's many ways to handle this, blocking the server ip-address to individual projects is one method, another is to block it in the hosts-file but since POGS uses ip-address instead of name for their scheduling-server I don't think this will work...

It's also possible to manually edit the <min_rpc_time> for a project in the client_state.xml-file located in the BOINC-data-directory, setting the deferred-time so long into the future it will report after race-start. This method is especially useful if you've got a slow internet-connection and the uploads and/or downloads are large, since it's only the scheduler-requests to the project reporting finished work (and asking for more work) being deferred, the uploads will continue as normal.

To edit, make sure BOINC-client is shut-down, open-up client_state.xml in notepad (or similar, don't use word or something). Search for your desired project, example http://pogs.theskynet.org/pogs and roughly 20 lines longer down you'll find the <min_rpc_time>

Chances are where's already a number here, something like 1398894615.588919
To set your desired time, remember this is seconds, so if you want to deferr for 24 hours just add 86400 to the number before the dot.

Alternatively, can use example http://www.epochconverter.com/ to choose your human-readable time, and get this converted to epoch-time. To wait until 04:00 UTC on the 5th. of May (meaning 4 hours after race-start), edit the line so it's something like <min_rpc_time>1399262400.123456</min_rpc_time>

 

[Assimilator1]

It is not like PrimeGrid races. You can crunch early but turn them in after the start and all will count.

If you get a bunch of WUs now, turn off network communication and then after the start send them in you will be ahead.

Trouble is: how do you get enough GPU WUs also? Do people use two clients to do this? One for pogs with communication turned off and another for GPU with communication turned on?

Never done a Prime grid race, but understood.

And I do crunch on my GPU (as you know), so I'd have to try RDs suggestion.
Although I have no idea what min_rpc is, or scheduler request.
Sounds like hassle, anyhow for now, it's late & I'm tired, I'll check it out tomorrow, time allowing.

 

[GLeeM]

To wait until 04:00 UTC on the 5th. of May (meaning 4 hours after race-start), edit the line so it's something like 1399262400.123456

Because I am 5 or 6 hours later I would add another two hours worth of seconds?

Did you have trouble getting enough WUs?
Did you just set the buffer to 5 days?

 

[Rattledagger]

Because I am 5 or 6 hours later I would add another two hours worth of seconds?

Hmm, since AFAIK the epoch-time is always UTC-time, I don't think it should be neccessary to add anything more... But, adding some more wouldn't be a problem, and you can always check in BOINC Manager afterwards if it's correct, since it's shown on the Project-tab how long a project is deferred.

Did you have trouble getting enough WUs?
Did you just set the buffer to 5 days?

I just used 10 + 10 days buffer, and started downloading after suspending all other projects. A few times did need to hit update, but appart for this just continued downloading work until started getting the "won't finish in time"-message. Taking a look, it seems the estimated run-times is shorter than the actual, so chances are will need to abort a little at the end. Keeping the buffer at 5 days and possibly even 4 days should probably be enough of a pre-cache, but it's possible an overclocked cpu will be faster than the estimates.


BTW, it's always recommended to wait a little before doing the initial dump, preferably everyone should check the start-stats for each project has been successfully recorded before reporting their 1st. results, since even the project doesn't have any problems it's possible the Pentathlon-stats is still problematic to collect.

 

[GLeeM]

It's also possible to manually edit the for a project in the client_state.xml-file located in the BOINC-data-directory, setting the deferred-time so long into the future it will report after race-start.

I got it to work on BOINC Manager 7.0.64. but it will not work with 7.2.42. The number just changes back or to zero!? Any ideas what might be wrong?

It continues to report WUs. And will not show "deferred" time.

 

[Rattledagger]

I got it to work on BOINC Manager 7.0.64. but it will not work with 7.2.42. The number just changes back or to zero!? Any ideas what might be wrong?

It continues to report WUs. And will not show "deferred" time.

Hmm, just tested with v7.3.15, it didn't work here either. So, looks like they've added some "safety-checks" on the numbers, if it's too long into the future it's ignored, or they've introduced a bug and the numbers aren't used any longer as they should be...

 

[Assimilator1]

Oh well, looks like a last minute switch afterall.

7.3? Is that beta?

 

[Ken g6]

For awhile, I was doing POGS work I'd downloaded and it wasn't flushing, so I thought I might have a large buffer for the race. But then it flushed without warning.

I do have a small backup plan, though. When I see a WU is ~85-90% complete, I'm suspending it. Once the race actually starts, I'll resume these and get a boost. Of course, this is very time-consuming for me to catch all these WUs. I think I'll suspend networking and not use my GPU tomorrow unless the GPU project is announced soon.

 

[Rattledagger]

7.3? Is that beta?

Alpha.

 

[Ken g6]

From the blog site:

The 3rd project of the BOINC Pentathlon in the discipline Natural Sciences is:
Rosetta@home (http://boinc.bakerlab.org/rosetta/)

All credits granted between
05/08/2014, 0.00 (UTC) and
05/13/2014, 0.00 (UTC) will be taken into the BOINC Pentathlon validation.

 

[Rattledagger]

The 2nd. project was just announced, this is for "Natural sciences"-disipline and the project is Rosetta@home located at http://boinc.bakerlab.org/rosetta/

The start is 08. May at 00:00 UTC, and the end is 13. May at 00:00 UTC.


Now, it's a very long time since I last ran Rosetta@home, but atleast one advantage is all wu's is single-redundancy meaning no wing-man. It's possible to specify the wu-times, with the shortest target being 1 hour and the longest 24 hours. Default is 3 hours. If not mis-remembers, some wu's doesn't work correctly with only 1 hour target, so possibly an idea to keep it at default... Note, if you're going directly to 24 hours, chances are most of the work will error-out due to "too long" cpu-time. Use incremental updates instead if wants longer wu's, and return some wu's with the new target before upgrading again.

Bandwidth-usage, where's commonly a large download (210 MB) but this is used for many wu's so possibly is enough with one large download. For uploads the longer the wu runs the larger the upload, atleast before some users with montly usage-caps did complain about large file-transfers so not sure if this has improved.

Disk-usage, atleast 300 MB/wu. Memory-usage, maybe the mini-rosetta-application is less demanding, but alteast the old Rosetta-application could easily use 500+ MB/wu, so for anyone with little memory this is one thing to look out for.

Deadline, 10 days, meaning the Rosetta@home-part of the Pentathlon finishes before the deadline is over.

 

[Assimilator1]

Wow R@H has big storage & RAM requirements!
Time to setup an account, been meaning to run this one for years! [edit] turns out I'd already setup an account 7yrs ago!, although for some odd reason I hadn't joined it to TA.

So are we focusing on 1 project or overlap as the projects do?

 

[Ken g6]

Well, I'm not doing Rosetta. I'm low on RAM as it is. Fortunately, POGS, the "marathon" project that runs through the entire race is a nice, low-RAM project.

 

[biodoc]

I've been building up WUs for skynet for more than 24 hours now and hopefully I'll have enough work in my que to last until the start of the race. All 4 of my rigs are in.

 

[Assimilator1]

How did you do that? Via the WU suspend method?

I would of done that but I forgot to stop getting new tasks for A@H until late last night, atm I'm just crunching those in the hope that I won't have too many that'll expire after I suspend A@H! :$ ......hmm I'd better switch the 4th core back to crunching. Bleh, GPU load is down to 60-70%, have seen worse before though.

 

[petrusbroder]

Three of my GPUs have a fan failure - each ...
Not much of a problem for the pentathlon, but still, ppd dropped somewhat terrible.

 

[Assimilator1]

3 in quick succession? Damn!