On Fri, 18 Jun 2010 10:40:28 +0200 Keld Simonsen <keld@xxxxxxxxxx> wrote: > On Fri, Jun 18, 2010 at 08:30:53AM +1000, Neil Brown wrote: > > > > 2-drive RAID10 f2 would be expected to provide better read throughput > > (possibly twice as fast) at some cost to write throughput. For many people > > this is a worthwhile trade-off. So it might be better for you. > > Read throughput would degraded down to write throughput (i.e. slower than > > RAID1) if the RAID10 were degraded. > > The slowdown of write thruput is on the raw raid, and on file systems > without elevator algorithms. If you employ a file system with an > elevator algorithm, then there will be no noticeable slowdown for > writes, as witnessed by many benchmarks. Filesystems don't implement the elevator algorithm. It is the driver for the disk drive which does that. So if you have a filesystem on RAID10 on 4 disk drives, then there are 4 instances of the elevator algorithm between the RAID10 and each disk drive. With RAID10 f2, consecutive chunks needs to be written to widely different locations on the device. The elevator certainly helps here by sorting the write requests so that all the even chunks go to one end of the device, then all the odd chunks go to the other end. The only difference that I can imagine between a filesystem and raw-device access is that maybe one allows the queue to fill up with more write requests before flushing them out, thus giving the elevator more opportunity to improve things. However I cannot see that difference in the code (though I might be missing it). One difference between writing to a raw device and writing to a filesystem is that when you close the raw device, all pending writes are flushed, while when you close a file the data can stay 'dirty' in memory for several more seconds. So writing to a raw device might seem slower because the close is slower. NeilBrown -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
