NetNews Usenet Archive 1992 #27/NN_1992_27.iso/spool/comp/benchmar/1764

Internet Message Format | 1992-11-23 | 5.8 KB

Xref: sparky comp.benchmarks:1764 comp.arch.storage:795
Newsgroups: comp.benchmarks,comp.arch.storage
Path: sparky!uunet!gatech!asuvax!ncar!csn!teal.csn.org!milton
From: [e-mail] (Milton Scritsmier)
Subject: Re: Disk performance issues, was IDE vs SCSI-2 using iozone
Message-ID: <[e-mail]>
Sender: [e-mail] (news)
Nntp-Posting-Host: teal.csn.org
Organization: Colorado SuperNet, Inc.
References: <[e-mail]> <[e-mail]> <[e-mail]>
Date: Mon, 23 Nov 1992 18:00:16 GMT
Lines: 100

In article <[e-mail]> [e-mail] (Dave Olson) writes:

> In <[e-mail]> [e-mail] (Milton Scritsmier) writes:
> | In its current drives, HP remaps entire tracks, even for one sector
> | errors (each zone has a certain number of spare tracks). This was
>
> That was not a great decision, if true.  I've not chased down
> just how HP does their bad block handling, as we haven't done
> serious eval of their drive (politics as much as anything else,
> even if it is a different division of HP from the workstations).

I was given a in-house technical document by HP about this design decision,
and from my experience with HP drives can verify that it is true from
having implemented the reassign command. (In this scheme consider the
possibility of having other bad sectors on the track that were not part of
the original request which did have a bad sector.)

However, I did not cite HP's design choice as being the best decision
possible about how to remap sectors (nowhere did I say that). Rather, I
only meant to point out that there is no method which is best for
performance, capacity, and other factors. Different drive vendors have
settled on different methods. If there was one obvious best choice, then
all drive vendors would have long ago settled on it.

> | a conscious design choice made after studying various methods. Some
> | of the reasons they gave were that media errors often overlap sector
> | boundaries, and that many disk operations often do not cross track
> | boundaries. In addition, head switching times to get to a slipped
>
> Sector slipping means that there *is* no head switch time, by definition
> (unless the sector slipped was the last one on the track, which is
> relatively rare).
>
> | sector pushed onto another track or to a spare at the end of a cylinder
> | often are several milliseconds. This is a significant fraction of a
> | seek, especially if you have to switch back to the original track to
> | pick up more data vs. seeking to a track where all the data is.
>
> Additionally, almost all SCSI drives offset sector numbers from
> track to track to allow for the head switch time.  There is
> still the latency issue, but at least you don't miss a rev.
>
> You *never* have to switch back to the original track if you
> sector slip.
>
> | No doubt sector slipping gives the best overall performance in general,
> | but with today's emphasis on capacity as well as performane, sector
> | slipping wastes too much media. In addition, the media has improved
> | to the point where a hundred errors on a 1 gig disk is high.
>
> You obviously don't understand sector wlipping.  It wastes no
> media at all.  100 errors on a 1GB disk is not at all uncommon.

Since you don't believe anything I say, how about looking at a couple
of facts:

1. As the number of defects per gigabyte becomes less, so too does the
   importance of the remapping method become less. And the number of
   defects per gigabyte *is* going down. At the company where I
   work we ship hundreds of 1.6 gig HP drives. Just to make sure that my
   feel for the average number of defects per drive was correct, I went
   back and looked at some hard copy defect lists I saved when I was
   testing samples of this drive taken at random from those we shipped
   (these are production HP drives, nothing special about them). One
   drive was truly horrible, and had 290 defects. Other sheets I have
   listed 69, 65, 41, 33, 33, 29, and 20 defects per drive. In addition,
   I remember one drive passing through here that had no defects at all
   (this surprised me so much I called HP and asked if a mistake had been
   made :-). Clearly the average number of defects per gigabyte on these
   drives is much less than 100 and that 100 defects per gigabtye is
   indeed high, as I originally said.

   I cite HP only because they are the drives with which I have had the
   most recent experience. No doubt other vendors' drives are in the same
   ballpark for similar capacities (I have one defect list for a Seagate
   Elite I which shows 12 defects). You can quibble about the actual
   numbers, but the number of defects per gigabyte *is* decreasing even as
   capacities increase.

2. I took a look at the drive specifications I have to see what various
   drive vendors do. I had six drive specs in all. HP uses the sectors
   per zone method in both its 5.25 and 3.5 inch drive families, while
   the Seagate Elite family and the Quantum ProDrive 1050S both use the
   sectors per cylinder method. In addition, from looking at the Format
   page of the Mode Select command it appears that the Maxtor MXT-1240S
   also uses the sectors per cylinder method (the "zone" used for
   describing the sparing parameters is a cylinder rather than a track,
   and the default number of spares per zone is less than the number of
   tracks per cylinder). The other two drive specs I had were for DEC
   and IBM drives. They were preliminary specs and did not provide any
   information on the method used for remapping sectors.

   Thus out of a sample of six well-known drive vendors, at least three
   of them (and probably more) did not use sector slipping for their
   latest drives. If sector slipping is such an obvious win, you
   might want to ask yourself why all these drive companies haven't
   caught onto it.