Date: Tue, 14 Jan 92 13:49:09 +0100 From: tthorn@daimi.aau.dk Which brings me to the question I've always wanted to ask: How do you use FFS on a SCSI disk? Given that SCSI disc's use logicial sectors and automatic bad sector remapping, it doesn't make much sence to talk of cylinder groups. Is it possible to avoid the logicial mapping? The short answer is: with great difficulty, and not without a lot of success. Usually, SCSI disks will generally keep logical sector whose block numbers which are close together in the same general area of the physical disk, but this is not guaranteed. Depending on how automatic bad sector remapping is done, the revectored blocks may be at the far end of the disk, or scattered through out the disk. So the FFS algorithms can be somewhat useful on SCSI disks by distributing files across the entire disk, and by reducing (somewhat) disk head activity by keeping files in the localized area wrt logical sector numbers. But of course it will not be as successful as when the kernel has a much better idea of where blocks will end up on the platters. Another thing which comes up with SCSI disks is that the elevator sort algorithm (which sorts the disk blocks in the request queue in an attempt to reduce head movement) will also not work as well as traditional disk controllers, again because there's no guarantee that the logical sector numbering has any relationship to the physical blocks (although there usually is some). However, there is usually some cache memory on the SCSI controllers on the disk drives themselves to help this situation a bit. Fortunately, the CPU time required to do the elevator sort and the FFS placement of disk blocks is small enough that it gets swamped by the time to do the disk I/O. So even if the elevator sort and the FFS don't help as much for SCSI disks, it shouldn't cost that much to use the algorithms in any case. - Ted
| Mark Lord | 2.6.25-rc8: FTP transfer errors |
| Kamalesh Babulal | Re: 2.6.23-rc6-mm1 |
| Greg Kroah-Hartman | [PATCH 025/196] paride: Convert from class_device to device for block/paride |
| Stephen Rothwell | Announce: Linux-next (Or Andrew's dream :-)) |
git: | |
| Linus Torvalds | Re: iptables very slow after commit 784544739a25c30637397ace5489eeb6e15d7d49 |
| David Miller | Re: [GIT]: Networking |
| Gerrit Renker | [PATCH 18/37] dccp: Support for Mandatory options |
| Jarek Poplawski | [PATCH] pkt_sched: Destroy gen estimators under rtnl_lock(). |
