Speed of Hardware RAID

One thing I have noticed recently, as result of the recent discussion on
hardware versus software RAID, is that the two drive LUNs we have created
through the Smart 5i hardware RAID in the DL380 G3 servers we have are
lightning fast by comparison to the same drives mounted JBOD on an Adaptect
39160. I'm not doing heavy I/O but just simple operations like opening the
drive and browsing folders, opening applications, etc. Is the 5i
controller just better implemented? Since I/O load is low I can't believe
the performance difference is attributable to the disk mirror being in
hardware. But the difference is quite noticeable.

--
Will

"Will" wrote in message
...
One thing I have noticed recently, as result of the recent discussion on
hardware versus software RAID, is that the two drive LUNs we have created
through the Smart 5i hardware RAID in the DL380 G3 servers we have are
lightning fast by comparison to the same drives mounted JBOD on an
Adaptect
39160. I'm not doing heavy I/O but just simple operations like opening
the
drive and browsing folders, opening applications, etc. Is the 5i
controller just better implemented? Since I/O load is low I can't
believe
the performance difference is attributable to the disk mirror being in
hardware. But the difference is quite noticeable.

--
Will



Will,

The biggest factor here is that the 5i is a cache-enabled, accelerated
controller. Also, you are distributing I/O across 2 (or X number of )
spindles, verses just 1 spindle with the JOBD config. When there is a
battery backup (BBWC) on the 5i, you get predictive read-ahead features not
offered by the Adaptec card. You can also set the caching policy for x%
read, x% write caching. There is also the matter of being able to select the
strip-size of your array volumes with the 5i, something you will never get
with that adaptec card. I would be neglectful if I didnt mention the awsome
driver implimentation that HP provides ....

So, there isnt just one thing that makes the 5i superior, although any 1
feature listed above makes it better by default.

Let me know if you have more questions,

- LC

"Nut Cracker" wrote in message
t...
The biggest factor here is that the 5i is a cache-enabled, accelerated
controller. Also, you are distributing I/O across 2 (or X number of )
spindles, verses just 1 spindle with the JOBD config. When there is a
battery backup (BBWC) on the 5i, you get predictive read-ahead features
not
offered by the Adaptec card. You can also set the caching policy for x%
read, x% write caching. There is also the matter of being able to select
the
strip-size of your array volumes with the 5i, something you will never get
with that adaptec card. I would be neglectful if I didnt mention the
awsome
driver implimentation that HP provides ....

So, there isnt just one thing that makes the 5i superior, although any 1
feature listed above makes it better by default.

I have the same volumes in software RAID, and there is effectively no I/O
load, so it cannot be the number of spindles that accounts for the
difference. Cache in system memory is faster than cache on the PCI-x
controller, so it cannot be cache.

The predictive read-ahead sounds really interesting, and I'm suspecting that
must be the feature that accounts for the perceived difference in simple I/O
read performance.

It's impressive.

--
Will

"Nut Cracker" wrote in message The
biggest factor here is that the 5i is a cache-enabled, accelerated
controller. Also, you are distributing I/O across 2 (or X number of )
spindles, verses just 1 spindle with the JOBD config. When there is a
battery backup (BBWC) on the 5i, you get predictive read-ahead features
not
offered by the Adaptec card. You can also set the caching policy for x%
read, x% write caching. There is also the matter of being able to select
the
strip-size of your array volumes with the 5i, something you will never get
with that adaptec card. I would be neglectful if I didnt mention the
awsome
driver implimentation that HP provides ....

So, there isnt just one thing that makes the 5i superior, although any 1
feature listed above makes it better by default.

Another follow-up: if in fact the performance difference is due to
predictive read-ahead technology, then one would assume that you get a very
high payback for keeping the hard drive defragmented at all times, since
read aheads would with certainty grab the rest of the data for any file you
were opening.

What is the cheapest way to turn a DL380 into an iSCSI server for a small
network? I'd like to compare the performance of using this machine as a
storage server using Windows NAS, iSCSI, and then compare that against a
low-cost NAS box. I suspect at gigabit speeds the network won't introduce
very much latency and we might actually see the benefit of that excellent
controller performance on the clients.

--
Will

I have the same volumes in software RAID, and there is effectively no I/O
load, so it cannot be the number of spindles that accounts for the
difference.

you are missing 2 huge points ... the stripe size you get with hardware raid
(and generally NOT with software raid), and the fact that software raid is
inherantly slower. With HW raid, you can read from 2 disks simultaneously.
With SW raid, you are reading from 1 disk only. The software parity engine
is used for writes only to keep the 2 disks in sync, and generally has to
issue 2 write commands for every data block. HW raid sends 1 write command
to the controller, and the partiy engine at the hardware level takes care of
the writes, not the OS.

Cache in system memory is faster than cache on the PCI-x controller, so it
cannot be cache.

Not the same type of cache usage. The cache on the array controller is used
for 2 things ... to store the read-ahead data (since disk cache is far
faster than accessing the data on the disk), and to store information that
needs to be written to the disk. In this manner, the OS sends a write
commands to the controller, which stores the data in the cache and at the
same time tells the OS that the data has been written to the disk. In this
way, the OS doesnt have to wait for the data to actually be written to the
disk, and this in effect 'speeds up' the operating system since it isnt
waiting for 'write complete' messages. It al happens in the background.


The predictive read-ahead sounds really interesting, and I'm suspecting
that
must be the feature that accounts for the perceived difference in simple
I/O
read performance.

Its definitely a factor, but I believe the improvement in system performance
is the aggregate of the many features I have been writing about.

You're impressive.

Why, yes. Yes I am. Thanks for saying so.

; )

--
Will

- LC

"Will" wrote in message
...
"Nut Cracker" wrote in message The
biggest factor here is that the 5i is a cache-enabled, accelerated
controller. Also, you are distributing I/O across 2 (or X number of )
spindles, verses just 1 spindle with the JOBD config. When there is a
battery backup (BBWC) on the 5i, you get predictive read-ahead features
not
offered by the Adaptec card. You can also set the caching policy for x%
read, x% write caching. There is also the matter of being able to select
the
strip-size of your array volumes with the 5i, something you will never
get
with that adaptec card. I would be neglectful if I didnt mention the
awsome
driver implimentation that HP provides ....

So, there isnt just one thing that makes the 5i superior, although any 1
feature listed above makes it better by default.

Another follow-up: if in fact the performance difference is due to
predictive read-ahead technology, then one would assume that you get a
very
high payback for keeping the hard drive defragmented at all times, since
read aheads would with certainty grab the rest of the data for any file
you
were opening.

ANY computer will run "better" when its file system is in its least
fragmented state. But, there is more to this than just running conventional
defrag tools and utilities. I would say that it is just as significant, if
not MORE so, to have an unfragmented MFT (for NTFS filesystems) as it is to
have defragemented files. The location on the disk of each file is stored in
the MFT, and for things like directory enumeration, a heavily fragmented MFT
translates into "slow" folder and file enumeration.


What is the cheapest way to turn a DL380 into an iSCSI server for a small
network? I'd like to compare the performance of using this machine as
a
storage server using Windows NAS, iSCSI, and then compare that against a
low-cost NAS box. I suspect at gigabit speeds the network won't
introduce
very much latency and we might actually see the benefit of that excellent
controller performance on the clients.


No idea ... havent done any work with iSCSI ...

But I would definitely look into your switches at the same time. GigE is
cool and all, but if you cannot leverage Jumbo frames (4k, 9k, 16k) then you
are missing out. Default is 1500 bytes (usually closer to 1472 or so, due to
packet structure overhead) ... im sure you have heard of the MTU. Well, its
related to that. Of course, this is much more useful for implimentations
where you re routinely moving LOTS of data around. Not ideal for a Domain
Controller, or DNS server.


--
Will

"NuTCrAcKeR" wrote in message
t...
You're impressive.

Why, yes. Yes I am. Thanks for saying so.

And your editing skills are very aggressive.

--
Will

"Will" wrote in message
...
"NuTCrAcKeR" wrote in message
t...
You're impressive.

Why, yes. Yes I am. Thanks for saying so.

And your editing skills are very aggressive.

--
Will


heheheh

Oh, come on. nobody can be 100% geek 100% of the time ... we have to take
our fun where we find it. However, the other material in the post is far
more worthy of commentary than the insignificant liberties taken in the
interest of self amusement.

=)

- LC