February 2001
DiskPerf delivers the
real deal on drives
Utility talks directly to IO subsystem
to profile disk performance
Review by Steve
Hammond
DiskPerf, where were you when
I needed you?
Actually, that may be
stretching it a bit, but our friends at Allegro have brought out a
tool that could be helpful to a site like mine that had a series of
disk failures just over a year ago.
That tool is DiskPerf, a
utility to track performance and IO capabilities for disk devices on
both MPE/iX and HP-UX systems. Why, you may ask, would one need such
a tool? Disk caching and virtual page storage may mask disk problems,
specifically performance issues. DiskPerf avoids those pitfalls by
talking directly to the IO subsystem, so it gives us the real
deal on the drives. Vendors will make claims, but DiskPerf can
determine the veracity of those claims. Need to find the slow drive
on your machine? DiskPerf will do that. How well is your disk array
performing? DiskPerf can help.
About a year ago,
my 987SX suffered through a series of hard drive failures. All these
failures were tied to the purchase of drives from a vendor whose
price was well below the other two bids we received. I am now quite
certain the drives were remanufactured although I do not believe they
were represented as such. If I had had DiskPerf following that
initial failure, I could have monitored the activity on those other
suspect drives. I might have seen degradation or poorer
performance on the drives before they failed over the following six
months. But I always said I wanted to spend a night sleeping on the
floor of my office...
DiskPerfs installation
is quite simple so simple, in fact, I could not remember how I
had done it and I had to call Allegro to ask them how I had done it!
Its just a restore followed by streaming a job. Thats it.
Demos and purchase copies are available on DAT or via Web
download.
The product itself is fairly
easy to run all command-line driven, with settings, INFO=
strings and even a startup file for regular runs. You can perform the
tests on a single drive, a subset of your drives, or all your drives.
Furthermore, you can run it with users online, since it performs
non-destructive reads to determine drive performance. It performs a
variety of tests on several levels to determine the performance of
the drives. I monitored my system with Glance while DiskPerf was
running and saw no overall performance hit, although my system was
not getting heavy user access at the time. DiskPerfs tests are
done to retrieve data randomly, to do multiple same page reads,
serial reads, serial reads within a 100Mb chunk of data on the disk,
reads where the next read is 1Mb higher than the previous read, and a
worst case scenario read (where the next read is the
furthest page from the current read).
If you tell DiskPerf simply
to test all, every one of these tests will be performed
on every disk on your system. As with any other Stan Sieler-created
product, the Help subsystem is extensive and very useful, so you can
find anything you need to know about tests or settings with ease.
(The Help system is extensive enough that there is no printed manual,
and I really saw no need for one.)
The final report
at the end of the process gives you the best and
average reads per second for both random and serial
access. It also provides the hardware path of the drive, something
which can come in handy, since you can never find that hard copy of
the configuration when you need it. For regular execution of the
program, you can either use INFO= strings in the run command, or you
can put the commands in STARTUP.DISKPERF.ALLEGRO, which it reads upon
execution.
As for my 987, DiskPerf told
me all the drives that I still have questions about seem to be
performing within expected tolerances, about within 10 percent of
each other on the various tests. Three of my drives are just within
the 10 percent factor of the other 15, so I will need to keep an eye
on those in the coming months.
One problem is that right now
in Stans words, DiskPerf needs a skilled operator to make
deductions. And this is true. You dont know what each
drive should be doing. What if they are all slower than the hardware
specifications? Youll just see they are within the 10 percent
range of each other, but that assumes they are working fairly close
to spec. That deductions problem is in the process of being resolved,
however.
Allegro has asked their
beta-testers and other HP e3000 developers who have received the
product to send back reports from their drives along with other
information on the devices. From this developing knowledge base,
DiskPerf will be able to start making some assumptions about the
performance of your drives. By determining the drive model and
comparing its performance to the knowledge base, DiskPerf will be
able to tell the user if a drive is really performing as it should.
This feature was not included in the version of DiskPerf that I
demoed, but it should be available as you read this.
Another interesting item is
the pricing of DiskPerf. The cost is $5,000, or you can rent it for
two weeks for $1,500. Allegro realized that this tool is not
necessarily something that one would need to run regularly, but could
be very useful occasionally when doing planning for system
upgrades and expansions or when attempting to diagnose performance
problems. It might be easier to sell your CFO on a one-time $1,500
expenditure with no support costs than the $5,000 hit with subsequent
support costs of $750 per year. You can rent DiskPerf for two weeks
and determine your problem and if you decide to buy it within
90 days, the rental price will be credited to your purchase.
After testing it, I would
describe DiskPerf as a tool that some of you will need all of the
time, and all of you will need some of the time. Allegro realizes
this, and has created a fine product with a pricing structure that
makes sense for everybody.
Steve Hammond is a system manager for
a trade association in Washington, D.C. who has been working with HP
3000s for 18 years and is chairman of the SIGPrint special interest
group.
|