Controlling Read-Ahead; Write Buffering - Alpha Microsystems AM-6000 Owner's Manual

Page
E-2
Programs that do significant random disk access (such as RNDRED) tend to slow down with this read-
ahead scheme. Most of the slowdown is caused by "thrashing" of the disk cache, where cache entries
that will be used again are removed from the cache due to the allocation requirements of the read-ahead
blocks (which typically are never used). The actual data transfer overhead is very little, as most SCSI
disk drives (especially fast SCSI-2 drives) have a track cache built into the drive allowing both the target
and read-ahead blocks to be transferred over the SCSI cable without delay.

Controlling Read-Ahead

For read-ahead to occur on the AM-176, both the AMOS disk cache (DCACHE.SYS) and the full SCSI
dispatcher (SCZR60.SYS) must be installed. The number of read-ahead blocks to be transferred into
cache on every physical disk read is contained in the disk driver. This can be set when using the
FIXLOG program to generate a disk driver with the appropriate read-ahead blocking factor.
When you use FIXLOG to create a driver for the AM-176 board, you are asked to specify the number of
read-ahead blocks. For example, type:
FIXLOG RETURN
FIXLOG.LIT Version x.x(xxx)
1. Change the number of logicals.
2. Create a sub-system driver.
Enter choice: 2
Enter name of generic driver to be used: SCZR60
Enter number of logical units per physical unit: 10
Enter SCSI id (0-6): 0
Enter number of read-ahead blocks (0-7): 5
Enter new driver name: DSK
New driver is now in memory.
To save the driver you have created, type:
SAVE DSK.DVR
If you wish to disable or change the number of read-ahead blocks, simply use FIXLOG to generate a
new disk driver. If the disk driver is for the DSK: device, don't forget to use MONGEN and embed the
new driver into the system monitor.
The generic AM-176 SCSI disk driver (SCZR60.DVR) is set up for seven read-ahead blocks.

WRITE BUFFERING

AMOS (and therefore every application written for AMOS) understands only 512-byte disk blocks.
Therefore, when a disk write request is made by a program, a single block transfer is made to the disk
drive. If the program then writes the next sequential block, the system must wait the latency time of the
drive (i.e., the time it takes the drive to complete one revolution) before the next block can be written.
Latency, even on fast SCSI-2 drives, is around 7ms.
RETURN
RETURN
RETURN
RETURN
RETURN
RETURN
AM-6000 Computer Owner's Manual, Rev. 00
Appendix E
RETURN