which approach to maximise speed with a SAN?

Hello,

I had a loong list of technical questions over the phone today for a
new job.
From How do you get the return code from a shellscript (easy - echo
$?) to Which approach to maximise speed with a SAN? I dunno, but as I
have to meet these guys for a face to face interview on Tues it might
be an idea to find out.
I can look into a comparison between RAID 1, 5 and 10 and some others
all by myself (cos i new about 1 and 5) and some others. But perhaps
someone on this list might help me out?

TIA

Tony

On 27 Feb 2004 15:17:42 -0800, (tony
barnwell) wrote:

Hello,

I had a loong list of technical questions over the phone today for a
new job.
From How do you get the return code from a shellscript (easy - echo
$?) to Which approach to maximise speed with a SAN? I dunno, but as I
have to meet these guys for a face to face interview on Tues it might
be an idea to find out.
I can look into a comparison between RAID 1, 5 and 10 and some others
all by myself (cos i new about 1 and 5) and some others. But perhaps
someone on this list might help me out?

TIA

Tony

Well, they've got some specific issue in mind I can tell you because
the question is infinitely open ended.

You can get a fast box like HDS or EMC (argh, I can't believe I said
that) and use raid 10. Raid 0 really won't make much difference due
to the huge amounts of cache except in the most extreme of
circumstances. But that's another area to look into.
Direct connect your server to a storage port, don't go through a
switch. I would doubt that you're bottleneck would be the switch
today but it's a possibility.
Get fast server hardware. Doesn't make sense to have 4 direct
connected fiber ports to an HDS when you're using an E250.
Multiple fiber connections. If redundancy is not an issue use the
standard 2 fiber connections as a single 200mb pipe. If redundancy
does matter then make it 4 ports.
If going through a switch put priorities on host traffic. Not all
switches do this so that's another consideration, switch performance.
McData and Cisco are fast, Cisco has more features and functions but
McData has a rock solid track record.

This plus about 30 more things can all be done to speed up SAN
performance for a single host. If it's the entire fabric they're
talking about, well that's another list. And at least as long...

~F

tony barnwell wrote:
$?) to Which approach to maximise speed with a SAN?

Like all performance issues, find the bottleneck and remove it, then
find the next bottleneck. You will need to understand the I/O load,
because the bottleneck will be at different places dependent upon the mix.

Factors to consider:
- Multipathing (i.e. PowerPath, DMP, etc) if the bottleneck is bandwidth
or queuing
- For highly random i/o's the memory centric enterprise boxes like
9900/DMX/Shark may not bring any performance above the midrange boxes
with a shorter path to disk (i.e. FasT, CX600, LSI-based controllers, ...)
- Physical layout, both in terms of raid level, striping, and how the
LUNs map to physical disks
- database and filesystem block sizes
- Host-side caching/memory issues like SGA size and buffer cache size
- Mapping between logical volumes on the host and underlying LUNs
- I/O contention within the SAN ( e.g. Inter Switch Links, contention on
the front end of the storage array, contention with tape traffic on the
HBA during backups )

Dave

dave dickerson wrote in message ...
tony barnwell wrote:
$?) to Which approach to maximise speed with a SAN?

Like all performance issues, find the bottleneck and remove it, then
find the next bottleneck. You will need to understand the I/O load,
because the bottleneck will be at different places dependent upon the mix.

Factors to consider:
- Multipathing (i.e. PowerPath, DMP, etc) if the bottleneck is bandwidth
or queuing
- For highly random i/o's the memory centric enterprise boxes like
9900/DMX/Shark may not bring any performance above the midrange boxes
with a shorter path to disk (i.e. FasT, CX600, LSI-based controllers, ...)
- Physical layout, both in terms of raid level, striping, and how the
LUNs map to physical disks
- database and filesystem block sizes
- Host-side caching/memory issues like SGA size and buffer cache size
- Mapping between logical volumes on the host and underlying LUNs
- I/O contention within the SAN ( e.g. Inter Switch Links, contention on
the front end of the storage array, contention with tape traffic on the
HBA during backups )

Dave
Thanx for both the responses. Comprendo. Getting the idea.
Had a look at some Macdata stuff.One comment related to "single 200mb
pipe" Should we be looking at Gb pipes ideally ie 5 times faster? And
possibly you might want to keep the redundant connections and avoid
single point of failure.

Liked the comment about iteratively find the bottleneck and fix it.
The interview question might have related to design of SAN in which
case you have to guess where the bottleneck might be. Should you maybe
start with the physical disks,layout (say, RAID10 for speed) and work
outwards towards the clients, (probably the most important clients
with be databases?) applying the guidelines above?

Thanks again

Tony

Thanx for both the responses. Comprendo. Getting the idea.
Had a look at some Macdata stuff.One comment related to "single 200mb
pipe" Should we be looking at Gb pipes ideally ie 5 times faster? And
possibly you might want to keep the redundant connections and avoid
single point of failure.

I hope he meant 2000mbit

Doing anything with 1gig equipment would seem like a bad idea since 2gig
isn't all that expensive any more.

Liked the comment about iteratively find the bottleneck and fix it.
The interview question might have related to design of SAN in which
case you have to guess where the bottleneck might be. Should you maybe
start with the physical disks,layout (say, RAID10 for speed) and work
outwards towards the clients, (probably the most important clients
with be databases?) applying the guidelines above?

If the storage is decent, one thing to look for is overloaded ISLs.

--
/Jesper Monsted

tony barnwell wrote:
dave dickerson wrote in message
...
tony barnwell wrote:
$?) to Which approach to maximise speed with a SAN?

Dave
Thanx for both the responses. Comprendo. Getting the idea.
Had a look at some Macdata stuff.One comment related to "single 200mb
pipe" Should we be looking at Gb pipes ideally ie 5 times faster? And
possibly you might want to keep the redundant connections and avoid
single point of failure.


Sounds like you may be confusing Mbits/Gbits with Mbytes/Gbytes in some of
the literature. The "single 200mb pipe" is referring to a 2Gbit link.


--
Nik Simpson