home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: biz.sco.opendesktop
- Path: sparky!uunet!looking!brad
- From: brad@clarinet.com (Brad Templeton)
- Subject: Re: SCSI-2 Adapters for SCO ODT
- Organization: ClariNet Communications Corp.
- Date: Sat, 02 Jan 1993 05:28:10 GMT
- Message-ID: <1993Jan02.052810.2589@clarinet.com>
- References: <1992Dec31.201533.15254@clarinet.com> <gXaqwB2w165w@zswamp.UUCP>
- Lines: 51
-
- In article <gXaqwB2w165w@zswamp.UUCP> geoff@zswamp.UUCP (Geoffrey Welsh) writes:
- > The questions are: how long does the transfer of a block monopolize the
- >SCSI bus (important when multithreaded I/O is taking place between the OS and
- >several SCSI devices) and how long the CPU must be shut down for the info to
- >be copied from the adapter into main memory.
-
- You're saying that DMA on intel machines shuts down the CPU? I wasn't
- aware of this. I thought DMA simply put the processor in a wait state
- if it happens to be going to main memory at the same time as a DMA
- transfer. Since the processor goes to main memory on fewer cycles than
- you would think (that's what the cache is for) DMA is not supposed to
- impact the processor a great deal if it is done right. This is not to
- say that EISA is not better than ISA, and that there isn't improvement.
- My goal here is to find out how much. It is important to realize that
- going to a 32 bit wide bus at 33 mhz is not going to give you 8 times
- as much througput as a 16 bit bus at 8 mhz. But what do you get?
-
- If your disks are sending at 1 mb/second, then you would have to have
- a lot of disks going at once before saturating a 5 mb/s SCSI-1 bus and
- finding yourself in need of SCSI-2. (Particularly since disks spend
- most of their time seeking and waiting for sectors to spin around)
-
- Don't get me wrong. I would love to have devices that are spewing out
- data at 10 mb/second so that I can use SCSI-2 and EISA speeds. But are
- these devices out there? I wouldn't mind even pushing the limits of
- SCSI-1, but I'm not convinced we're doing that. (Other devices, such
- as tapes, CD-roms etc. aren't anywhere near the usage levels of the
- disks.)
-
- If a disk stores 35 sectors/track, uses only one head at a time and spins
- at 3600 rpm -- this describes most disks -- then 1 megabyte/second is
- all it can sustain. A surprisingly easy fix that's not done more is to
- read from multiple heads at the same time. On a 16 head disk you can of
- course get 2 bytes per bit-time, 16 mb/second, though you become more
- vulnerable to errors. Anybody doing this?
-
- Another method is to double to 7200 rpm, which has the advantage of reducing
- rotational latency to 8ms -- or increasing density.
-
- One scheme larger systems use is to have several disk drives and store one
- bit per disk drive! With 19 disk drives, you can store 16 bits and 3 bits
- of error correction. This makes it possible to take any one disk drive
- offline -- and the system still works. Indeed, when a drive goes bad on
- these systems, you replace it and slowly rebuild the ecc onto it. You
- are vulnerable to errors while it is off and being rebuilt, and you must of
- course keep the drives in sync.
-
- Doubt anybody is doing this for SCO, but it's a way to get highly reliable
- disk operation and 16 megabyte/second throughput.
- --
- Brad Templeton, ClariNet Communications Corp. -- Sunnyvale, CA 408/296-0366
-