Who does better striping, disk tower or O.S ?

[Matthew Ghofrani]

I am running 11.11 on HP_UX rp8400s & my disks folks insist that the tower (in my case EMC) does a far better job striping disks using a 1 Meg stripe chunk compared to my way which is using o.s. striping with a factor of 64 or 128 kbs. I do have apps ranging from using 20 Gigs to 500 Gigs. Also, my understanding is when os sees for example 10 disks presented to it vs. one disk, it dispatches more agents to handle things. Any thoughts ?

[Florian Heigl (new acc)]

I also think Your should definitely do the striping inside the EMC^2. I'll try to give some reasons:

- tagged command queuing will show more benefit
- write cache of emc^2 can better adjust to Your accesses
- Your LUNs are already spread over multiple disk spindles inside the EMC^2, so You'll possible stripe on top of a stripe set.
- ease of administration, i.e. when extending lvols or with mirror setups.

practically You should benchmark both scenarios, but I think Your storage people are right.

[Devender Khatana]

Hi,

Stripping at hardware level is always a better option. As it reduces overhead from OS and system and they are available for handling other transanctions and better utilize your storage tower.

It does not really effet to have one or ten LUNs as ultimate I/O is happenning to same box. It is rather better to have multiple LUNs and then have them accessed through different controllers. This will increase I/O performance. But the no. of LUN's should not be very high for the ease of managibility.

HTH,
Devender

[Matthew Ghofrani]

ok so I see the need to elaborate. My physical chunks in the tower are 8.5 Gig each
If I ask for 85 Gig, they would use 10 disks on 10 different spindle, make 1 meta disks=85 gig & present it to me or they would give me 10 disks on 10 different spindles with each disk=8.5 gig and I would make my VG and LVs striping across 10 disks. So, overall in both cases we end up using 10 different spindles while I use 64 kbs chunks & vs. theirs of 1 Meg and supposedly in my case O.S. dispatches 10 times as much agents to handle IOs. Am I right ?

[Alzhy]

Let's not forget that OS Striping (via LVM or VxVM) complements and further improves the performance of already striped (or RAIDed) storage sets (LUNS) from your "towers" (arrays).

With the exception of EVAs, practically all others arrays (EMC, Hitachi, HP's XP line, etc.) will still require you to stripe the storage objects (LUNS) that are already striped (or RAIDed) in hardware inside your array. The reason for OS striping is simple - you get to "build" OS based storage units (i.e. LVOLS or VOLUMES) that will engage a larger number of physical disks inside your array - hence will result in better performance and no hotspot.

So there you go.. OS striping can be treated independent of your array's internal striping or RAIDing (as some stripes are actually RAID5 sets inside your array).

[Alzhy · Accepted solution]

And btw, your Disk Folks are wrong if they think you need not do any more striping on the host end. Whether you're using LVM or VxVM, you can get the most out of your EMC LUNs that are already striped (RAID-S or RAUD10/5?) by further striping them on the host end.

[Steven E. Protter]

In Keenen's book HP-UX CSA, as a result form his vast experience, he recommends that striping and mirroring be handled on the hardware level.

That means the disk array.

Doing it with software means the cpu and resources on the hp-ux box have to be expended to do the job and there are better uses for this resource.

In summary: better to do it on the hardware.

SEP

[Alzhy]

SEP, et al.

I think there is some level of misunderstanding here. The HW striping (be it RAID0,S,10 or RAID5) that are used to present (or carve) an array is different than the OS level striping (via VxVM or LVM).

There will always be disk architectures that without added host based striping - one can't get the most out of the array. EMC and Hitachi based architectures cannot present a striped or RAIDed LUN to a host that engages a large number of physical disks inside the array. Therefore, one will need still OS based striping to "assemble" these already striped/RAIDed storage units into larger and higher performing storage sets (volumes/filesystems) so it engages more physical disks inside the array.

[Matthew Ghofrani]

I guess the jury is not out on this yet. How about 10 times agents dispatched from OS vs. one in my previous example? Does that have any merit in comparrison to the os overhead?