New version seems to improve performance

[dmike]

Has anyone noticed this? I just recently downloaded the latest version of BOINC, having been running the release prior to the latest for the last week or so, and I have noticed that CUDA tasks complete about 60 seconds faster.

Is it possible that the new BOINC software can have an affect on crunch times, or does the BOINC application have little or nothing to do with the processing of WU themselves?

Running Einstein btw.

[pinhodecarlos]

I will test the new version as soon as I clear my cache.

[dmike]

Cool, let me know.

Also, I did notice a performance increase when putting BOINC on a SSD as opposed to a traditional HDD. Roughly 45-60 seconds per WU faster on average.

I've chatted with folks on the Einstein@Home forum and the project administrator said that the app being on an SSD could not possibly increase performance, but it did for me.

[Mumps]

Cool, let me know.

Also, I did notice a performance increase when putting BOINC on a SSD as opposed to a traditional HDD. Roughly 45-60 seconds per WU faster on average.

I've chatted with folks on the Einstein@Home forum and the project administrator said that the app being on an SSD could not possibly increase performance, but it did for me.

You talking about GPU WU's? Or just CPU WU's? One thing to keep in mind is that BOINC is constantly re-writing the client_state.xml file. Every time a new WU is downloaded, or a completed one uploaded. Re-write. And even as various in progress WU's checkpoint, it re-writes that file. So, if your system was indeed I/O bound, the SSD could have meant a slight improvement. So one other thing that could affect performance similarly is simply changing the frequency BOINC is allowed to check point. The default is every 60 seconds for each task it's running. Upping that to a higher value may provide a bit of improvement, at the expense of losing a bit of work if the system is a laptop or something that gets rebooted somewhat frequently. My preferences I have that set to 300 seconds, or 5 minutes, rather than the default 1 minute. But you could raise that to whatever your tolerance level deems is an appropriate amount of acceptable loss. Some projects actually honor it even. :)

And, if your system was memory constrained, but you let BOINC gobble up everything it wanted anyways, accessing the swapfile could pretty easily mean that kind of improvement in times moving to an SSD drive.

[dmike]

Thanks for the insight. I'll be upping the check point time.

This particular box that I'm referring to has been on 24/7 since the first time I turned it on. It has never locked and only has rebooted after certain software installs. Other than that, it is on 24/7/365.

[pinhodecarlos]

I still didn't update Boinc to the latest version because I don't know if a reboot is needed. If so I can only reboot my machine next Sunday due to a post-processing job that I am running for NFS@Home.
Meanwhile on Collatz I raised the checkpoint time from 60 seconds to 1000 seconds and didn't notice any kind of decrease of wu's crunching time.

Carlos

[Rattledagger]

You talking about GPU WU's? Or just CPU WU's? One thing to keep in mind is that BOINC is constantly re-writing the client_state.xml file. Every time a new WU is downloaded, or a completed one uploaded. Re-write. And even as various in progress WU's checkpoint, it re-writes that file.
The checkpointing was moved to individual files, if not mis-remembers for the v6.10.xx-clients, so this won't be a bottleneck any longer, except if an application writes very large checkpoints on it's own. All the other things on the other hand does impact things, and the larger client_state.xml is, the slower things goes. The size depends both on #wu's and #results, since results takes more space than unstarted wu's anyone building-up a large cache of results during the Pentathlon will get ever increasing client_state.xml until starts reporting the results.

BTW, even if disk-I/O doesn't become a bottleneck, just for the BOINC-client to generate an updated client_state.xml will use some cpu-resources, and the larger the file the more cpu-resources will be wasted by the BOINC-client.

[pinhodecarlos]

A few hours ago I updated BOINC to the latest version (7.0.64) and I can't notice any kind of improvement. Some wu's take longer some wu's take less time. With a slow GPU I can't have a universe of wu's to statistically say that I get a decrease in wu processing time of X seconds.

Carlos