Hi,
I posted a question recently on the practical difference in "speed"
(response time, througput, etc.). for a 10ms vs. 15ms hard disk which
contains a heavily used NFS file system. I received a quite a few
responses which are listed below. Here is an additional piece of
information from the book "System Performance Tuning" by Mike
Loukides:
"In many applications, perhaps in most, the seek time is the most
important indicators of a disk's performance---even more important
than the raw throughput... The ratio of time spent seeking to time
spent transferring data is usually 10 to 1, and often much higher.
Therefore, if you reduce the seek time, you will see a big performance
improvement." (pg 135).
The response below seem to confirm this assesment.
Thanks to all who responded to my query.
- Henrik
>From warrane.connect.com.au!ups!uniq.com.au!kevin Sat Sep 11 01:14:09 1993
Seek time is probably the single most important factor in that
scenario. Take a look at the papers by Auspex on cheap SCSI disks
vs. IPI in USENIX and other proceedings. The conclustions that seek
time dominates xfer speed are still true, even with the faster SCSI
busses.
>From cmspringer@amoco.com Sat Sep 11 03:04:33 1993
I would suspect that some of the disk performance will be absorbed by
the network, but I have these ideas to offer:
We use Seagate Wren9 1.9GB formatted drives for database usage in
development. They are very durable and are cheap for their capacity.
(approx $1400 each). For production, we use Seagate Elite II 2.1 GB
drives which cost more but perform noticably better. both are 5.25"
form factor.
If you are concerned about performance and need something approx. 2
GIG, I highly recommend the Seagate Elite II. For 3.5" form factor I
recommend the Seagate 1.05GB unit. But, in this unit there's one at
11 ms and one at 14. the 14ms unit's part number ends in K.
>From ilkka@tks.oulu.fi Sat Sep 11 03:55:53 1993
get the fast ones. i have made some tests and the write speed
correlates almost one to one with the speed of the disks. with reads
the buffer cache can compensate a lot if you have enough memory. with
nfs-mounted partitions the speed of the server has less influence
until you use all of the power of the server. currently we are using
seagate st11200 and st42400 disks in suns without any problems. sun
knows them by name sun1.05 and sun2.1. i havent tried yet the still
faster seagate barragudas.
>From bacon@mtu.edu Sat Sep 11 09:51:52 1993
well, think about it this way. a 10ms disk will be roughly 1/3 faster
than a 15ms disk, right?
Another point: Most all of the new wunderbar 3.5" disks that are in
the 10ms range also spin faster (4500-5400rpm), reducing disk latency
and increasing the data rates on/off the platter as well.
There's also fast SCSI-2 to consider, but if your server doesn't handle
that, or if you tend to do a lot of small transactions, I don't believe
that this is so necessarily important, so long as you don't put over a
couple drives on any given scsibus.
NFS traffic tends to be very random - different requests from
different clients, rather hard to attempt to do read-ahead on,
requiring a lot of mem for cacheing. On top of that, NFS writes are
synchronous, so anytime you write, you are heavily dependent upon the
speed of the disk.
I just recently got my hands on a bunch of the segate st11200N 1GB
wunderbar drives. The difference is incredible. They can move data
probably twice as fast as the old Wren-VI or Wren-VIII class drives I
used to rely on (5.25", ~15ms, 3600rpm). (I've done a certain amount
of testing with these drives - I don't have precise figures or graphs,
I just hand you that number as a generality based on various practical
observations of mine in the "real world". But one of those ST11200Ns
will handily sustain 90 disk transactions/sec all day. You'd be lucky
to get 60 off a Wren-class drive, 45 is more realistic.) There's no
excuse not to buy these drives, esp since mass quantity pricing has
kicked in now, and you can buy these 1+GB 3.5" drives for near $1/MB.
If you really want k**k-a** drives, check out those new seagate
barracudas. ~2GB, 9ms, 7200RPM(!). Reasonably cheap too. I want
one. Badly.
>From bathurst@noaapmel.gov Sat Sep 11 20:40:31 1993
The RPM of the drive is really the big differing factor.
My experience is 20% between a 5,400 RPM drive and a 3600 RPM drive.
A Seagate Barracude (7200 RPM) would be about 40% faster.
>From vasey@issi.com Sun Sep 12 00:14:59 1993
We have many types of fast and slow disks in use, and don't notice
much difference anywhere, but the file type and processor mix is so
varied, it's hard to pin down how much variance is due to the disk
access time.
I suspect the key phrase above is "over the network". Since this is
your biggest bottleneck anyway, superfast access time may not be too
critical; however, since a bottleneck exists already, that is also a
good reason to remove anything else that could occasionally degrade
server performance. And since the price difference for disks that
have ALL the latest technology, including speed (8-11ms) and
reliability (300-500K hours) is pretty small these days, I'd vote for
faster (and hopefully better) drives. Good luck!
>From dannyd@mel4360.miden.com.au Sun Sep 12 19:11:14 1993
Hi Henrik,~
do you want simple answer ??? 50% is difference.Sorry , it is not so
simple. If you want to be sure go for faster disk , but you have to
pay much more. On the other hand you can use slower disk with out any
problems. It is depend on the application , which you are running. If
you are running any database ( a lot read/write), you will be better
of with fast disk( A customer had ORACLE database and search took
nearly one hour - in this case - better was fast disk). If users will
be students , you can get nicely with slow disk. You still have other
option , to increased size of buffers in the kernel and other tuning
can be done.
>From bern@kleopatra.uni-trier.de Mon Sep 13 03:29:04 1993
No practical Experience with 10 vs. 15 ms (or any other Jump) Disks,
but if you're in for this Kind of Optimization, you should rather
consider:
-- Buying differential SCSI Equipment (installed mine today, 10 mb/s
on the Bus instead of 4 mb/s with ye olde SCSI Bus)
-- Buying very special wide SCSI Equipment (no Experiences)
-- Buying some Cacheing Product (e.g., PrestoServe SBus Card; wiped
our HDs once we didn't know what we were doing ...)
-- Distribute the Files over several Devices or, even better, Buses
to distribute Execution Wait Times (You may use SUN's DiskSuite
(if I didn't misinterpret the Ads) to distribute the Data without
splitting up Directories/Devices on the Software Side)
>From shandelm@jpmorgan.com Mon Sep 13 06:58:40 1993
In your environment a fast seek time is critical. You have no way to
predict how, where and when the users will consume disk space. After
time the disk heads will be bouncing from one edge of the disk to the
other servicing I/O requests from the different users. A fast seek
time will help give back some performance when this happens.
>From djiracek@jupiter.fnbc.com Mon Sep 13 07:31:02 1993
I have nothing but experience to back up my opinion, but I would say
that you are not going to see any difference. What you need to do is
to make sure you have ample swap space and memory so that the data is
cached effectively. I would like to see a summary on this one.
>From netcom.com!highserv!dallas.highland.com!kevin Mon Sep 13 10:45:46 1993
Henrik -
Access time is important, but there are also other factors that will
increase performance: latency, cache size, caching algorythms, etc.
There are many examples of drives with higher access times (15ms)
outperforming lower access times (10ms) due to better latency and
cache setups.
>From schara@inf.puc-rio.br Tue Sep 14 09:31:20 1993
The problem with speed is always this: your system will be as fast as
the slowest link. Depending on your configuration, the bottleneck
could be on bus throughput, network throughput, cpu speed...
So you may not see any difference with the more expensive disk (or not
enough to make you buy it). A server with a lot of memory is always
better than a fast disk (but that is just my opinion).
Also, disk usage depends on software and also system configuration.
Small local disks (for swapping) made greater difference on our
network than a large fast disk.
>From warrane.connect.com.au!ups!uniq.com.au!glenn Mon Sep 13 18:32:08 1993
Heaps of difference. Always go for the disk with the fastest seek
times as this is more critical then transfer rate. This is especially
for multiuser and nfs where the heads will be randomly moving all over
the place all the time.
-- Henrik Schmiediche, Dept. of Statistics, Texas A&M, College Station, TX 77843 E-mail: henrik@stat.tamu.edu | Tel: (409) 845-9447 | Fax: (409) 845-3144
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:08:16 CDT