On Monday 15 December 2014 18:32:14, Irgendeiner wrote: > Am 15.12.2014, 17:28 Uhr, Diego Roversi <[email protected]> wrote: > > On Sat, 13 Dec 2014 12:29:10 +0100 > > > > Irgendeiner <[email protected]> wrote: > >> Hooked to a BananaPI board, reading speeds are above 100 MB/s and > >> follow the speeds limited by the disk drive. However, writing is > >> about 4 times slower and I could not yet find an authorative > >> answer explaining it.> > > It can be a problem of partition not aligned to 4096 byte block? > > > > https://wiki.archlinux.org/index.php/Advanced_Format > > > > On my hardware (olinuxino-a20-micro) writing is slighty slower, > > 45Mb/s vs 60Mb/s. > > > > Check also if you are using a frequency governor different from > > performance. > > The partitions are perfectly aligned, otherwise the high reading > speed could not be obtained.
The read-ahead cache could hide the mis-alignment during read. But I can more or less reproduce your results... > The governor could only be the culprit when for writing it would not > step up or not use both cpu. I will check that again, but imho > there must be some deeply hidden cause. > > In your case when you get reading speeds of only 60 MB/s there must > be some other bottleneck and this most probably hides the slowdown > of writing speed. > > Any other hints? > I am pretty sure that everybody using BananaPI with fast enough 3.5" > disk is suffering from that! ... On a cubietruck with a Samsung SSD 830 128GB, I get: read 101 MB/s write 41 MB/s Not factor 4 but definitely a big difference. When reading, the cpu freq goes from 720000 to around 912000. When writing, it stays at 720000. I don't think this accounts for the observed difference. Kernel is a 3.4 sunxi variant. It may be interesting to try this with a mainline >= 3.16. -- You received this message because you are subscribed to the Google Groups "linux-sunxi" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
