Re: fr1 speed calculation tricks...
"Peter T. Breuer" <[email protected]>
| Newsgroups | gmane.linux.enbd.general |
|---|---|
| Message-ID | <[email protected]> |
"Also sprach denis bonnenfant:" > First, I set speed_limit_min and max to respectively 1 000 000 and 2 000 000. > this way, i can test the real speed of the link during reconstruction. > I forced full reconstruction of nda, and got a speed of 38 000 KB/s, with a > system load of 8! Well, the load is just the number of processes waiting in i/o, so that probably means 4 daemons, or servers, or something like that. The "real speed" will probably be different from what you think, because on write enbd can skip writing blocks that are/will be the same both sides. Measuring read and write separately may show you the effect. > So I set min to 1000, hoping that md will go down to this limit. No effect ! It's a minimum - when raid sees the speed is below that it increases the priority of the daemons, so as to bring the speed up again, hopefully. > I set max to 10 000, this time it works, speed fall to this value, and load fall > to 1. Setting min value seemed to have no effect on real reconstruction speed, > weird! I don't think you'd expect it to. You can look at the throttle code in md.c to see how it works. > Then i try the same test with an intelligent resync of a clean disk : > > set min to 1000, max to 10000. Reconstruction speed falls to 1000 ! I raise max > to 1 000 000, no effect. I raise min to 1 000 0000, speed goes to this large > value ! That's right. > So it looks like > - max is corresponding to the maximum speed to be set for true reconstruction, That's what it means! The raid code will insert a wait when speed goes above it. That will likely slow the resync down :-). > - min is corresponding to the minimum speed for intelligent resync. > So min must be set much higher than max Well, yes. Priority of resync thread is increased when speed falls below here. > Am I right ? Well, it's not really a mystery! The throttle loop is in md.c. Peter