On Thu, Dec 17, 2009 at 03:54:54PM +0100, Robert wrote: > David Gwynne wrote: >> let's not wave our hands around in the air for no good reason. surely there >> are more important things to get upset about. > > I have to object here. > > If I understood the original article correct this would mean that the > writing of 1 sector could in fact be 2 reads + 1 write (internally in > the disk), which would be a *serious* performance hit (we are not > talking 5-10% here...).
Really? got numbers to back that up? I'll bet money there is no noticeable performance loss in a real world load. Maybe a slight drop in some sort of stupid micro benchmark. Let me give you a hint, the drive spins really fast and to read the data of the platters it uses heuristics to reconstruct the data. These heuristics require multiple passes of the head over the sector that is being retrieved; this is leveraged by the read-ahead cache too. Another thing to consider is that 512b is the quanta, not the actual value used in an OS. You are not going to find a very significant number of 512b IOs; it is much more likely to have >2k and more likely > 4k IOs. The performance impact we are debating here is only for not rounded IOs, it is a numbers game and I don't see how this would affect either negatively or positively the daily use of a disk. Highly specialized applications could benefit but lets count how many of those use WDC SATA drives. I can see the BS already coming our way, OMG MOAR bytes!!! totally bettar!!! > I don't know about the effect on RAID controllers that you mentioned, > but I guess their cache helps to hide those effects and also they are > not in use on every PC. And drives don't have cache? > The affected SATA disks on the other hand will be in use on basically > every new i386/amd64 system in the near future and I don't know how much > their internal cache will help. I see a big problem coming up... ? I don't unless we decide to go change every piece of code that makes the 512b assumption.
