Memory use doubles after 4.5 minutes

tbzep
tbzep
Joined: 18 Jan 17
Posts: 2
Credit: 137790757
RAC: 0
Topic 224967

I've been trying to solve an issue with Boinc/Einstein.  Nearly two weeks ago, Boinc/Einstein started showing 100% disk usage and locked the computer up.  I ran task manager and didn't see a change in memory usage.  I later realized task manager was lagging. I started using Rainmeter to track memory and disk usage because it was reacting faster.  With it, I saw that memory was jumping to 100%. Evidently it is hitting the swap file hard and locking up the computer.   The odd thing is that this doesn't happen until about 4.5 minutes after I start crunching work units.  I changed nothing on the computer when it started happening.

Win10 Pro (work computer so I can't add memory)  

BOINC 7.16.11

8gb ram

plenty of free space on HD, auto sizing of pagefile.

I run all 4 cores (8 threads) and the ATI GPU at 100% CPU and 100% memory when the computer isn't in use.  I've done it this way for years, first with Seti, now with Einstein.

I understand that the Gravitational Wave Search units may take more memory.  They are the only work units that are on the computer right now.  I don't understand why it chugs away on all 9 work units with no problem for 4.5 minutes, with the progress percentage clicking away showing good progress, then the memory usage suddenly maxes out.  

To try to solve the issue thinking it was a disk/swap file problem I have:

Run chkdsk /f /r

Put in an exception on AVG to make sure it doesn't scan anything in the Boinc folders

Checked drivers

Updated Windows

With advice from the main Boinc message board, I changed it from using all 4 cores to a single core and GPU.   I'm running 3 work units right now.  That allowed me to see the doubling of memory use at 4.5 minutes without dipping into the swapfile.  It is using as much memory with the three work units as it did two weeks ago with all 9 work units going at the same time. 

 

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5842
Credit: 109406434596
RAC: 35375508

tbzep wrote:....   The odd

tbzep wrote:
....   The odd thing is that this doesn't happen until about 4.5 minutes after I start crunching work units.

It's known behaviour for GW GPU tasks to significantly increase the memory in use at some point after the starting phase.  It's been pointed out a few times in discussions over the last year or more.  I don't run any CPU tasks but it may well occur with CPU tasks as well.  The main focus of people giving reports has been on GPU crunching.  A lot of data needs to reside in memory and, apparently, this progressively builds up in the early stages to reach a peak well after the start.

You don't give details of your GPU but looking at your Windows machine, it's listed as an R5 350X.  The Wikipedia GPU database lists it as a GCN 1st generation card and fairly low end.

It has been reported in a number of different messages at Einstein that GCN 1st gen cards cannot handle the GW GPU app.  I've tested it myself and know it's the case.  All tasks eventually error out with a "TIME_LIMIT_EXCEEDED" error message.  If you look at your GW GPU tasks on the website, you will see that all have failed and none have succeeded.  A lot were canceled since they hadn't started by deadline.  If you click the TaskID link for any that did attempt to crunch, you will find an exit status like the following:-

Exit status:      197 (0x000000C5) EXIT_TIME_LIMIT_EXCEEDED

The only type of GPU task that you can crunch with that GPU will be the gamma-ray pulsar (GRP) ones.  You should stop (in your preferences) using the GW GPU search and abort all those that you have on that machine.  Change your preferences for that machine to use the GRP (FGRPB1G) GPU search instead.  You can use different 'locations' to have different preferences for your two machines.

Your Windows machine has 4 physical cores (8 threads) so if you try to run 8 CPU tasks and a GPU task as well, you are going to get poor crunch times and laggy behaviour.  Try running just a single GRP GPU task with all CPU tasks suspended.  This will give you an optimum crunching time for that type of task.  Then (with those tasks still running) start unsuspending CPU tasks as needed to get an idea of the performance when crunching 1 or 2 or 3 or ... (etc) CPU tasks in combination with the GPU task.  Don't be in a hurry - be sure the results are repeatable.  You will easily be able to pick the optimum number of concurrent CPU tasks to run with the GPU task that will give you the best output.  When you find that number, set your preferences to give you that combination.

tbzep wrote:
I run all 4 cores (8 threads) and the ATI GPU at 100% CPU and 100% memory when the computer isn't in use.

You should find that less than 9 concurrent tasks will give more output in a situation like yours.  Take time to do the experiments properly.

Cheers,
Gary.

tbzep
tbzep
Joined: 18 Jan 17
Posts: 2
Credit: 137790757
RAC: 0

Thanks, Gary. From my

Thanks, Gary.

From my understanding, even if I can get another 8GB of ram, it's still going to have issues with Gravity Wave work units because the current video card just won't handle it.  It might eliminate the swapfile issue, but I'll still end up with bad work units.  I doubt they will let me swap video cards, but I will check on it.  I know some of the computers here have nVidia cards, but I don't know which models they have in them and I don't know if they have any spares floating around.  

I've never realized I can choose the type of work units. I've unchecked all of the Gravity Wave work unit types.  Unless/until I can make a hardware change on this work computer, I will just run the various pulsar search units.   

The home Linux computer has much better hardware and seems to be fine with whatever is thrown at it. Now that I've had this swapfile issue, it's caused me to come here and poke around on the forums and look at completed work unit data.  I wish I had done so long before I started having the issues.  It looks like this computer has been wasting cycles all this time.  

Thanks again

tbzep

Comment viewing options

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