XP12000 Storage - SAN or Direct Connect?

[Alzhy]

This may seem to be a Storage Issue but it actually has some LVM/VxVM Implication. Here it goes:

Knowing that the XP Storage line can have LUN presented to more than one front end port (also called "CHIP" or Client Host Interface Processor) - Is there any rationale/advantage to using a SAN implementation at all? This ability to present LUN(s) on more than one front end port seem to defeat the need for a SAN .. specially since the XP12000 can have as many as 128 (or is it 256) front end ports.

Why?

Say on a direct connect configuration I have LUNS presented to a prod server. If I do mirror breaks - I can simply present the mirror LUNS onto a backup server or possibly a dss/test/dev server that are also direct attached to the XP array - so there really is no need for a SAN anymore - and I think simpler. On the XP/HDS line of arrays - it is possible to "present" or "path" a LUN to more than one front end FC connection.

Obviously this configuration will be high performing as well - but is this a wise plan.. ? Will it be more expensive than having a few front end ports and just rely on the SAN for more server connectivity?
Will Arry CHIP ports be more expensive than SAN switch ports?

[Steven E. Protter]

If you are the only one going to be using data on this array, then there is probably no need for the san.

The san provides you the ability to have more servers than there are ports on the front(or back) of the array to the same shared data.

If there is no need for lots of machines to access the data on this array then eliminated the san accomplishes the following:

1) Decreases complexity of a san.
2) Lowers the overhead involved with a san switch.
3) Elimnates one potential point of failure
4) Probably marginally improves performance.

I'm sure there are other reasons.

I am assuming that just as with our disk array the interface directly on the array is compatible with the fc card on your server.

None of this makes direct connect worth it if there is even a possibility of other servers needing to access this array. Reconfiguring access is a pain and not worth it all.

Question: Have you considered MC/ServiceGuard and Alternat links in your configuration plan?

Regads to Harley man.

SEP

[Alzhy]

The environment will consist of probably 12 servers with 2-8 FC HBA's -- or a total of I'd say no more than 64 FC's. While it may be an overkill to provide All 64 FC front-end ports from the XP12000 - will cost savings in establishing a SAN be enough to justify the extra $$ to have the one-to-one Array port to Server HBA count?

The rational SAN functions as I've already menioned can be done by the XP's LUN multi-pathing features.. And apart from this feature - I really could not think of any compelling reason for a SAN. Regarding tape backups -- we will have a dedicated TAN (Tape Area Network) and will just do simple via-the-system bus backups on a backup server (disk to system bus to SAN to tapes). We will not be doing any Serverless backups (ie. moving blocks from Disk to Tape via the SAN)


SEP.. the Harley is in the freezer... Come Spring.. I'll probably take it to Chicago..

[Duncan Edmonstone]

Without checking, I'd be willing to bet that the cost per port for a CHIP pair is higher than the cost per port for a typical FC switch - there's one reason for using a SAN... the other reason is to achieve over-subscription - the average server doesn't pull anywhere near the amount of IO that a CHIP port can generate, but 2, 3 or 10 servers could - thats a better use of available resources.

Of course, if your environment is especially IO intensive across only a small number of large servers, then go withiut the SAN...

HTH

Duncan

[Steven E. Protter]

Answer to question 2.

Probably not.

The 64 server situation would probably better be served by a san with a couple of redundant brocade switches to provide for alternate links and serviceguard failover/sharing.

SEP

[Alzhy]

Our environments are VERY io intensive.. Certain batches needing sustained I/O bandwidth of up to 20K IOPS and up to 400 MBytes per second.

SEP.. we'll be using VxVM so no PVLINKS here.. just VxVM's DMP features.

Our switch infrastructure is core-switch based so the modules themselves will I think be pricey as well but Duncan you're probably right - it will be on orders of magnitude more expensive than the switch ports.

[Peter Mattei · Accepted solution]

Well, a single XP12000 port can do up to 47000 IOPs (512Byte blocks) and up to 181MB/s (256kByte blocks) so it takes large servers to make them busy.

Currently you can have up to 224 FC ports on a single XP12000.
Why 224?
You have 8 slotpairs for either CHIPs or ACPs.
Up to 7 can be CHIPs and up to 4 can be ACPs.
So 7 CHIPs with 32ports each = 224 ports.

The benefit of a SAN in your environment would be cost savings! (XP ports are more expensive than switch ports)
To maintain performance you have the possibility to share a big number of ports between servers and thus allow a statistic load balancing.

Think of having 8 HBAs per your 12 servers.
You can either buy 96 XP ports and connect directly or fewer ports and add switches.
With switches you would share groups of XP ports between groups of servers and DMP would balance your load and maintain equal load on all ports.

As an HP employee I love it if customers buy lots of ports but as a technical consultant I do not want to oversell!

So at the end it is physics. Add your average and peak Servers IO load and compare it with the XP performance maximums. This will help you determining the right configuration.


Cheers
Peter

[Vincent Fleming]

I've seen lots of customers direct connect to XPs - when you need the performance of dedicated ports, nothing beats a direct connection. Better yet, you can use High-Performance mode; it'll halve the number of ports you can use (ie: you buy 32, but use only 16), but will significantly improve random performance.

Also, using Veritas DMP will balance load from any one server, but will not evenly distribute the load across the frame's ports, unless you map every LUN to every port, and use switches to allow every server access to every port. Of course, you can do variations on this, such as present 1/4 of the LUNs on 1/4 of the ports, so you end up with 4 groups of servers hitting 4 groups of ports. (don't forget that there is a maximum # of luns on a port - so this type of configuration can significantly limit the number of LUNs you can support).

As for what's cheaper - it depends, actually. Enterprise switches (ie: McData-style) can be expensive in low numbers of ports, as you're paying a lot just for the box. They get more cost-effective with lots of ports.

So, for pricing, have your Storage SA work up configurations with switches and without, and compare for yourself. You just need to know how many total ports you'll need. ie: you have 12 servers; you want 4 HBAs per server; that's 48 ports. It's possible that you can get 48 ports on CHIP pairs cheaper than 16 ports (1 CHIP) and 4 16-port switches...

Much of all this really depends on your budget. If you can afford the Rolls Royce, why settle for a Kia? (assuming that there really is a big difference in price in your configuration)

Direct-connect is the way to go with a small number of servers, if you can afford to do it, and need the performance. If you go with switches, I can guarantee that you will get more latency, and more maintenance - no switches means that the switches can't break.

Cheers,

Vince

[Peter Mattei]

Vince, you have to be careful what XP you are referring to!
High performance mode is only available on XP512.

On an XP12000 things are different! You need to know that a single CPU services two ports. If you want to get the highest possible performance out of a single FC CHIP ports on XP12000 only use half of the ports (no setting required)

Cheers
Peter