Re: ccd performance (was: [ GEOM tests ] vinum drives lost)
: a lot of information on how to use it. I generally recommend : using a stripe size of 1152 for multitasking loads. : :Sectors? Why particularly this value? It's described in 'tuning'. Basically you want a fairly large stripe to reduce multi-disk seeking when reading sequential files (that is, if you do not need the combined bandwidth of more then one drive for the sequential case), and you also want to use a stripe that does not cause meta-data (e.g. inodes and bitmaps) to wind up on just one drive, e.g. use 1152 instead of 1024. or you will wind up with unbalanced accesses. :ccd, but I know a lot of cheap hardware RAID arrays always read an :entire stripe at a time, which requires more memory and takes longer. :Have you checked ccd for this? : :Greg I've done extensive work on ccd. It does not try to read a whole stripe, it just breaks the I/O up as appropriate, issues duel-I/O for mirror writes, and tries to select a reasonable (single) side when doing a read from a mirrored area. I even have a little code in there to try to reduce unnecessary seeking when reading from a mirrored area. But, again, CCD is not trying to implement 'real' RAID. It can't rebuild a lost mirror drive, for example, and does not implement RAID-5. IMHO A real RAID controller with NVRAM should be used for those things. -Matt To Unsubscribe: send mail to [EMAIL PROTECTED] with unsubscribe freebsd-current in the body of the message
Re: ccd performance (was: [ GEOM tests ] vinum drives lost)
Matthew Dillon wrote: But, again, CCD is not trying to implement 'real' RAID. It can't rebuild a lost mirror drive, for example, and does not implement RAID-5. IMHO A real RAID controller with NVRAM should be used for those things. FWIW, the people who sell RAID controllers with NVRAM feel the same way about software RAID implementations... -- Terry To Unsubscribe: send mail to [EMAIL PROTECTED] with unsubscribe freebsd-current in the body of the message
ccd performance (was: [ GEOM tests ] vinum drives lost)
On Sunday, 6 October 2002 at 11:30:16 -0700, Matthew Dillon wrote: Yes, ccd is fairly light weight. 'man tuning' and 'man ccd' has a lot of information on how to use it. I generally recommend using a stripe size of 1152 for multitasking loads. Sectors? Why particularly this value? Only use a small/tiny stripe size if you need single-tasking sequential performance (and even then you can take tune the stripe to the drive's own caching capability). The biggest mistake most people make when using striping is that they use too small a stripe size which causes nearly every read() or write() to have to be split across multiple drives, which multiplies the overhead, or causes sequential reads of medium sized files to constantly seek multiple drives, destroying the effectiveness of having two seekable heads in the first place. Pretty much exactly what I preach. One disadvantage of large stripes is that they require careful coding to optimize. I haven't looked at ccd, but I know a lot of cheap hardware RAID arrays always read an entire stripe at a time, which requires more memory and takes longer. Have you checked ccd for this? Greg -- See complete headers for address and phone numbers To Unsubscribe: send mail to [EMAIL PROTECTED] with unsubscribe freebsd-current in the body of the message
