n x EVA Equals One Hitachi/HP XP1024?

[Alzhy]

How many EVA must one need to reach the scalability and performance of an XP1024?

Is it 2 or 4? Or Is it the other way around -- how many XP1024's to reach the "performance" of an EVA5K?


Also, would it be advisable to do software based Stripes across LUNs on say 4/8 EVA's? Will still use SecurePath for Path Protection.. Just need more IO bandwidth and IO/s...

[Ted Buis]

Nelson,

From a capacity standpoint it takes 4 EVAs (240 drives per EVA) to basically match the potential of the XP1024. From a cache perspective, the maximum cache on the XP is many many times greater than the EVA. The XP has the potential for many more CHIPs (Channel Host Interface Processors ) or front end controllers. However on a per LUN basis, the EVA can be much faster because it can internally stripe across many physical drives, while on the XP you really need to stripe to get max performance. As always with performance it depends on the application usage. The EVA is much easier to manage and much easier to get good performance, but if you have to do OS striping anyway, then I would likely go to the XP. The XP has the availability edge over the EVA, but also comes with a price tag to match. The EVA is much more modular and is a true virtual array and so can may snapshots. Do you need to do CA for data replication between units? If so, the XP has the edge. Just remember on both arrays, the number of spindles controls the back end performance, so this can be important. There is a 8c8d version of the EVA5000 for highest performance (8 contoller, 8 drive bay or really 4 arrays), so it is reasonable to expand the number of controllers.

[Peter Mattei]

Some more facts
Max LUNs per box: EVA = 256, XP1024 = 8192
Max # of CA Pairs: EVA = 128, XP1024 = 8192
Backend performance: EVA = 30kIOPS or 675MB/sec, XP1024 = 57kIOPs or 2000MB/sec
max # of FC ports: EVA = 4, XP1024 = 64
# of OSes supported is much higher on the XPs.

The XP has the possibility of reaching a higher sustained performance when the load is well distributed.
On a single LUN basis the EVA will get higher performance due to internal striping. On the XP an LDEV (usually=LUN) is sptiped across 4 or 8 disks. On the EVA a LUN is striped over between 8 and 240 disks!
This is good for OSes like Windows, Novell etc. where no real good Volume MAnager is implemented!

Cheers
Peter

[Alzhy]

Thanks for the thoughts/technicals...

On the issue of scalability, I am finding out that the EVA really cannot scale well due to its 4 Frontends. It only behaves/performs well if there are a few number of servers connected to it. The impact of one server doing intense I/O is very noticeable on the other connected servers.

We're thinking of striping (host-based) accross LUNS on 4 EVA's to increase performance and lessen WaitIOs on our UNIX boxen. What do you think? Bad/Good? Or ditch the EVA for the more predictable Hitachi (XP) -- as one HP Account Rep once suggested...?

[Peter Mattei]

Well, if you stripe across 4 EVAs you will probably get more performance than with a single XP1024.
The disadvantages are:
- Availability. Think of an EVA at round 99.99% availability. 4x EVA striped = 99.96%. An XP1024 means 99.999% availability
- Continuous Access (Remote Copy). If you plan to use CA you will get into trouble. 4 EVAs cannot be synchronized over CA. Thus data consistency cannot be achieved in an emergency (if the primary DC goes down you will most probably loose the replication links at different times, thus the 4 EVAs at the second DC are out of synch)

The main reason for the XP1024 is not only performance but availability, scalability and the rich feature set.

Cheers
Peter

[Peter Mattei]

PS: What about assigning points?

[Mike Naime · Accepted solution]

Nelson:

In order to really compare Apples to Apples, you need to put some additional fact behind what you are trying to connect to the storage arrays.

1.) What is your fan-out ratio? I.E. For each fibre port. How many servers are you attaching to it?

2.) How much storage per server do you need? Keeping question #1 In mind. The XP1024 offers you more storage ports to handle the load of the storage array than the EVA does.

3.) What OS are you connecting? The EVA was designed for VMS and TRU-64 Unix. The XP1024 grew out of a UNIX world. The EVA is best for an Alphaserver. The XP1024 is best for the Unix systems. While you can opperate on the other platform, there are connectivity issues that you will have to deal with.

4.) If you are looking at raw storage capacity of the storage arrays.... Kiss performance goodbye!!!


If you really want performance out of your EVA5K. (Or any Storage controller for that matter) Buy lots of small disks to populate it with. 240 36GB disks will give you the best performance on the EVA 5K and will force you not to overload the system with tons of servers because "I have space left"

If you put 146GB disks in there. Management will keep loading it up until you run out of storage space. And then ask you why the performance sucks. They don't like hearing "I told you so!" :-)

If you load up the XP1024 with 150TB of disk space, and limit it to 4 fibre ports, you will have performance issues too!

Mike Naime

[Rob Young_4]

Hmmm... Now how do I reference or see the
post above? Guess I open a separate window.
This ain't Usenet.

Peter Mattei wrote:

The XP has the possibility of reaching a higher sustained performance when the load is well distributed.
On a single LUN basis the EVA will get higher performance due to internal striping. On the XP an LDEV (usually=LUN) is sptiped across 4 or 8 disks. On the EVA a LUN is striped over between 8 and 240 disks!


This isn't correct. This topic was discussed on comp.arch.storage a while back
and a good contribution describing XP's LDEVs
is found here:

http://tinyurl.com/6tpme

I'm not trying to displace the EVA completely, everything has its place in
your situation the EVA may be the better of the two. If I _had_ to place
this on the HDS I'd most likely be using Open-E-CVS(1000MB) LDEV's spread
across 19 Raid Groups. which would give me the access to 76 disks. I may
even consider to lower the LDEV size and use 190 Open-E-CVS-CM(100MB)
spread across all of the Raid Groups giving me access to 760 disks. Which
should yeild better performance than the 165 disks in the EVA. Assuming
you are using some form of software striping (LVM, etc) you can stripe
accross all of these LDEV's to gain better performance. Of course if I
had an 18GB database with the requirements you do I'd load the whole thing
into memory and place it on the cheaper EVA disk.

---

So essentially you could spread a LUN across
760 drives in the XP1024. Feature-wise, I
think it may be one of the strongest storage
arrays out there. Of course there is a certain vendor that would trumpet their superior cache bandwidth, but you certainly
don't have the flexibility of the XP. How
many EVAs to equal an XP1024? Capacity-wise
that is easy math. Performance-wise? I
thought at one time EVA could state a good
case on some workloads with 4 EVAs. But
if you had tremendous write traffic - the
EVA is limited to 170 MByte/sec when in
mirrored-cache mode - it is no contest. Stating a case for very high IOPS, I thought the EVA won as Peter did. But I learned you
can hit 760+ disks with a Hitachi LUN (maybe
all 1024 disks but the convuluted example I posed was 18 GByte LUN and as you can see
Jake used 760 disks to create that hypothetical LUN).

Rob

[Peter Mattei]

Well Rob

Sorry, you are not quite right and there are some points to add!
I am an HP storage consultant designing XP solutions since 1999 when the XP256 came out and am a big fan of the XP architecture. I just want to correct!

It is true:
You can create a LUSE (not CVS) of up to 36 LDEVs an spreads it over 36 Array Groups .... but LUSE do not stripe in small junks but concatenate.
Concatenating means you fill the first LDEV and then go to the next LDEV.
And, HP does not recommend building LUSE with > 8 LDEVs.
LUSE gives you larger LUNs but not performance. You want to use that for Windows systems where you cannot really use LVM to build a volume group of multiple LUNs to get great performance (like on HP-UX)!

Still I believe that the XP1024 is the most advanced and highest available array on the planet but you have to know how to configure and use it!

Cheers
Peter

[Rob Young_4]

Peter,


"On the XP an LDEV (usually=LUN) is sptiped across 4 or 8 disks."

As you point out, you can create LUSEs up to 36 LDEVs but your best
practices is 4-8 LDEVs but at 8 LDEVs they would be touching 32 disks
not 4 or 8. So then there is the whole
issue of concatenation and broken OSes having
an issue with that.


"It is true:
You can create a LUSE (not CVS) of up to 36 LDEVs an spreads it over 36 Array
Groups .... but LUSE do not stripe in small junks but concatenate.
Concatenating means you fill the first LDEV and then go to the next LDEV.
And, HP does not recommend building LUSE with > 8 LDEVs."

Okay.

"LUSE gives you larger LUNs but not performance. You want to use that for
Windows systems where you cannot really use LVM to build a volume group of
multiple LUNs to get great performance (like on HP-UX)!"

Windows no LVM? You can stripe Dynamic Disks. Now maybe you are clustered - splurge on Veritas. But we are outside Nelson's
scope. Nelson talks about performance and mentions Unix boxes and "We're thinking of striping (host-based)".

That being the case, he can serve up those nasty old concatenated LDEVs
and stripe a filesystem across them. He could literally touch hundreds
of drives and multiple fibre ports and the concatenated LDEVs won't be
a performance issue.

So circling back to his question, how many EVAs to make a XP1024 in
performance? I think it could be quite a few... and maybe you could
build a pathological case for Nelson where EVA outperforms it, but I don't think
so - not unless you have quite a few EVAs.

I think Unix with a good LVM makes up for
weaknesses in storage back-ends. There are
OSes of course that EVA helps out a lot (not a good host based striping capability or limited host based striping capability).

Rob