man.bsd.lv manual page server

The Interleave Factor

If a ccd is interleaved correctly, a “striping” effect is achieved, which can increase sequential read/write performance. The interleave factor is expressed in units of DEV_BSIZE (usually 512 bytes). For large writes, the optimum interleave factor is typically the size of a track, while for large reads, it is about a quarter of a track. (Note that this changes greatly depending on the number and speed of disks.) For instance, with eight 7,200 RPM drives on two Fast-Wide SCSI buses, this translates to about 128 for writes and 32 for reads. A larger interleave tends to work better when the disk is taking a multitasking load by localizing the file I/O from any given process onto a single disk. You lose sequential performance when you do this, but sequential performance is not usually an issue with a multitasking load.

An interleave factor must be specified when using a mirroring configuration, even when you have only two disks (i.e. the layout winds up being the same no matter what the interleave factor). The interleave factor will determine how I/O is broken up, however, and a value 128 or greater is recommended.

CCD has an option for a parity disk, but does not currently implement it.

The best performance is achieved if all component disks have the same geometry and size. Optimum striping cannot occur with different disk types.

For random-access oriented workloads, such as news servers, a larger interleave factor (e.g., 65,536) is more desirable. Note that there isn't much ccd can do to speed up applications that are seek-time limited. Larger interleave factors will at least reduce the chance of having to seek two disk-heads to read one directory or a file.

Disk Mirroring

You can configure the ccd to “mirror” any even number of disks. See ccdconfig(8) for how to specify the necessary flags. For example, if you have a ccd configuration specifying four disks, the first two disks will be mirrored with the second two disks. A write will be run to both sides of the mirror. A read will be run to either side of the mirror depending on what the driver believes to be most optimal. If the read fails, the driver will automatically attempt to read the same sector from the other side of the mirror. Currently ccd uses a dual seek zone model to optimize reads for a multi-tasking load rather than a sequential load.

In an event of a disk failure, you can use dd(1) to recover the failed disk.

Note that a one-disk ccd is not the same as the original partition. In particular, this means if you have a filesystem on a two-disk mirrored ccd and one of the disks fail, you cannot mount and use the remaining partition as itself; you have to configure it as a one-disk ccd. You cannot replace a disk in a mirrored ccd partition without first backing up the partition, then replacing the disk, then restoring the partition.