I use Boinc Tasks, and recently I added column with Virtual Memory size for tasks. Today I noticed that on one hosts WUProp uses 90 GB of memory - that's a lot! On other hosts usually this is below 100MB. What is more interesting, this particular host runs 4 CPU and 5 NCI tasks only. Look like some bug for me.

Screenshot is here: https://pasteboard.co/GBxrcJu.png

---

Wow thats a lot, especially for so few tasks running.

There was a user on my team that said he turned off WUProp and his MW times dropped a noticeable amount and he gained 400k RAC across several computers. (Several 7990s). He was going to turn it back on after a day or two to see if MW times went back up.

That got me to check out the CPU time on some of my hosts and noticed one went up when I switched from Asteroids over to Rosetta and then to WCG
http://wuprop.boinc-af.org/results.php?hostid=122461

The tasks with CPU time 200+ seconds was with Rosetta, the others with Asteroids and WCG. Same number of tasks running. I didn't check memory usage.

Does the WUProp memory/CPU usage vary depending on how long a CPU task has been running?

I have collected coredump from running process (90GB file). Now I am compressing it, xz should do good job here. Seb, you will have a lot of data to analyse :)

Edit: 90GB compressed to 14MB, nice :). Seb, I sent you PM with download link and system details.

Edit2: This is link to this WU: http://wuprop.boinc-af.org/result.php?resultid=72876366. It is still running now, should complete in about 2.5h.

---

Yesterday's WU finished successfully. Today I have another one with 90GB of memory allocated on this machine. Something fishy is going there. I suspect massive memory leak.

---

..."90GB of memory" does not sound right. I see you are running Linux on a 32-core PC. The Windows version of Wuprop does not report back about memory usage on the task detail page.

I also use BoincTasks to monitor my multiple PC's & tasks. The largest WuProp task I currently have running is using aprox 77MB (MegaBytes) Virtual Memory and aprox 7.7MB Real/Physical Memory on a 4-core Windows 8.1 PC.

---

..."90GB of memory" does not sound right. I see you are running Linux on a 32-core PC. The Windows version of Wuprop does not report back about memory usage on the task detail page.

I also use BoincTasks to monitor my multiple PC's & tasks. The largest WuProp task I currently have running is using aprox 77MB (MegaBytes) Virtual Memory and aprox 7.7MB Real/Physical Memory on a 4-core Windows 8.1 PC.

Please check again. I have one Windows machine, and both BoincTasks and BoincMgr reports virtual memory size for WUProp - in my case about 10MB.

---

That is interesting. I did a quick check of my systems using Process Explorer and found for most hosts the virtual memory size was ~50-60MB. However I did find on one of my hosts it was 33GB. I closed BOINC and relaunched it. The virtual size was still 33GB for the data collect app. I restarted the system size it had been up for about 45 days & the virtual memory size was still 33GB for the data collect app on the host.
It's not really a problem, just a little odd.

It is strange. Today memory usage dropped to 38MB. What is more interesting this happened not only for this one host, it look like it dropped on other too. I am sure about one other, which was using 1GB - I do not see this number anymore. Other ones also looks smaller to me. I checked app list and I do not see any new app version added there recently, so most probably Seb changed something on server side (probably params passed to app on start). I suspect he recently changed something to get more data and ban more cheaters. Seb, could you tell us more what was going on recently?

---

I didn't make any changes on the server.

The problem occurs only on linux hosts with NVIDIA GPU.

I think it is not a memory leak. The CUDA driver registers all the GPU(s) memory + host memory in a single virtual address space using the kernel's virtual memory system. It isn't actually memory consumption, per se, it is just a "trick" to map all the available address spaces into a linear virtual space for unified addressing.

---

I didn't make any changes on the server.

The problem occurs only on linux hosts with NVIDIA GPU.

I think it is not a memory leak. The CUDA driver registers all the GPU(s) memory + host memory in a single virtual address space using the kernel's virtual memory system. It isn't actually memory consumption, per se, it is just a "trick" to map all the available address spaces into a linear virtual space for unified addressing.

I saw some CUDA threads in coredump, so it might be. However this does not explain why this memory consumption suddenly dropped to reasonable level. Also that machine where WUProp used 1GB is mysterious, it was either my Windows box or one of ARM devices without Nvidia GPU. I will monitor this memory consumption to see if it will go up again or stay low.

---

I didn't make any changes on the server.

The problem occurs only on linux hosts with NVIDIA GPU.

I think it is not a memory leak. The CUDA driver registers all the GPU(s) memory + host memory in a single virtual address space using the kernel's virtual memory system. It isn't actually memory consumption, per se, it is just a "trick" to map all the available address spaces into a linear virtual space for unified addressing.

I saw some CUDA threads in coredump, so it might be. However this does not explain why this memory consumption suddenly dropped to reasonable level. Also that machine where WUProp used 1GB is mysterious, it was either my Windows box or one of ARM devices without Nvidia GPU. I will monitor this memory consumption to see if it will go up again or stay low.

My Windows 7 host has stayed consistent at 33GB and I just checked the notebook I use for work, which I didn't check previously, and it is reporting 2TB in Windows 10.
Since my other machines with Radeon or Intel GPUs and are not exhibiting this behavior it does seem likely to be related to Nvidia GPUs or maybe some implementations of OpenCL.