Checkpoint time setting

[Foxter]

Hello

Could you please tell me how to set the checkpoint interval in the 8.4.9 client version?

In the old 7.x version there was a slider between 3 minutes and half an hour in the client config. I cannot find that setting in the 8.4.9 version.

Since a I have a machine that folds 24/7 with very seldom disturbances from other activities, I would like to set the checkpoint at 30 minutes (maximum interval) to maximize the eficiency of the folding.

[Joe_H]

That checkpoint setting only changed the frequency for CPU folding. Current GPU folding cores and WUs use settings made by the researchers which are so many steps through the total. Most are set to amounts ranging from 2 to 5 percent of the progress. The client can not change those, and the frequency is important as that is when some data is collected and sanity checks are run.

If you could change the CPU WU checkpoint frequency it would not make any significant difference. The default was 15 minutes, tests made in the past showed almost no difference in processing time between that and 30 minutes. The real differences showed up with those who used very low settings for the checkpoint interval like 3 or 5 minutes. It may have made more difference when processors were much slower and a WU could take more than a day or two to complete.

[Foxter]

Thank you for the clarification.

[arisu]

For GPU cores, this is tracked in issue #247. It may be a while before anything like that is implemented for GPU cores. For CPU cores, it would be quite easy but the performance improvement would be negligible because the checkpoints are very fast anyway.

The best way to speed up CPU checkpoints would be to fix whatever the core is doing when it saves files, because it does this really weird, sloppy, and pointless fcntl() and lseek() spam as well as a bunch of fsync() calls that aren't even performed at the right time (i.e. more data is written after that call and it's not synced a second time!). That alone would decrease the checkpoint overhead. Although the overhead isn't very severe to begin with.

[muziqaz]

CPU checkpoints, even at default values, are very fine grained. Unless you are restarting every 5 seconds, you usually resume at the point where you paused

[arisu]

Yeah they are very fine grained. I think they will stop at most 25 steps past the interrupt signal, which takes no more than a second or two, and they resume from exactly the step that it stopped.