[Date Prev][Date Next] [Thread Prev][Thread Next]
[Thread Index]
[Date Index]
[Author Index]
RE: Big drives and Mylex?
- From: "Kirk R. Erichsen" <kirke inteli-home com>
- To: axp-list redhat com
- Subject: RE: Big drives and Mylex?
- Date: Mon, 13 Dec 1999 08:56:30 -0800
Even the "Smart" Array 3200 is little more than a Compaq OEM of the DAC960
PG with 2.x class firmware.
It is for this reason that many consultants like myself may be tempted to
flash the EEPROM up to a higher Mylex "retail" version when there is a
second EEPROM seated in the second EEPROM chip seat. Of course, this will
end your warranty when done, though in certain very specific circumstances
(or when you are at the end of your warranty), this can get you support for
the 18.2, 24 and even 36 GB drive capacities that you may require. Of
course, remember the old water cooler anaology when building RAID arrays:
the more spigots you have on your water cooler the faster you can drain (or
fill) the water cooler. Having a REALLY LARGE water cooler with only a few
spigots gets you a high capacity per cm but not much in the way of
read/write performance. Somewhere in the middle is the right
price/performance/capacity which can vary wildly with system or applicaiton
requirements. It is still preferable to build an array with 9GB building
blocks (particularly the newer IBM Enterprise or Seagate Cheetah 2 drives).
As a simple example, I have found that for large sequential read/write and
random uncached read/write, the fastest RAID level by far is RAID "6" (in
Mylex speak), also called RAID level 0/1 by the RAB (RAID Advisory Board).
RAID level 0/1, sometimes called RAID "10" is the RAID level in which you
have two identical RAID 0 stripped arrays (consisting of two or more drives
per RAID level 0 array), which is by definition a "mirrored stripe set"
which is very different from a stripped mirror (you can research this at
the RAB, if you wish). At this RAID level, you do not use parity
information which means there is no real performance hit when a drive is
lost since parity isn't being used to generate missing data and there is an
ENTIRE copy of the data stripped across the other array. The effect of
RAID 10 is high performance through multiplexing two RAID 0 arrays (RAID 0
has a high read/write performance due to stripped data across multiple
disks with no parity generation overhead). A properly configured RAID 10
can be as much as 50% faster than a RAID 0 array of the same size, though
you need twice as many disks to achieve the same capacity.
Large stripe block sizes (say 128 KB) produce the best sequential
read/write performance by "chunking" more data per write (and therefore
fewer transactions are required) while smaller stripe blocks (say 8KB) are
ideal for random access to such things as databases. I turn write-back
cache ON on the controller in all installations since the UPS that is
installed is always more than enough for a 15 minute run, more than long
enough to shutdown the system and commit to disk. I don't bother stuffing
more than about 32MB of memory (which is pretty cheap these days) since the
performance difference between 8 MB of cache and 32 MB in most cases is
less than 6%.
Where sequential read/write of large files is the design for the RAID, use
read-ahead caching. Most new RAID controllers have Adaptive cache
configurations in which the cache can reconfigure itself into several
smaller caches in order to take better advantage of the available cache for
smaller more random traffic. I am also pretty hellbent on getting even
more performance by ENABLING the write cache on the drives themselves. By
default, most hard drives ship with their mode pages for write cache
DISABLED to make them "safer." If you are using a newer drive with a
large, smart cache, take advantage of it in the the critical area of write
performance and turn them on. Note you will see NO performance increase in
situations where the ratio of read to write time is greater than 75% (which
is most of the time) but you will see a slight (5%) difference if your RAID
spends most of its time being hit by a steady stream of write requests (say
in log generation for a database with a long sequential stream of data
writes).
StorageWorks kits with a duplexer (adds channels to the stock backplane)
can produce the ideal result though you may be able to buy a less expensive
unit from the industrial computing market with similar or even greater
capacity.
Good luck,
--Kirk
> Per Compaq's docs and config tools, the Mylex controllers will not
>support drives in excess of 9GB - in other words, Compaq will not sell 36GB
>drives spec'd into a system with the Myles RAID controllers. I'm trying to
>set up RAID using 36GB drives and was wondering why they are not supported
>by Compaq on that controller.
> This is the only viable solution I can use, as we are a Compaq
>reseller and need to demo a "fully supported" config that can be ordered
>direct.
>
>thanx
>Don
>
>--
>To unsubscribe: send e-mail to axp-list-request@redhat.com with
>'unsubscribe' as the subject. Do not send it to axp-list@redhat.com
>
>
[Date Prev][Date Next] [Thread Prev][Thread Next]
[Thread Index]
[Date Index]
[Author Index]
[]