EVA Sparing and VRaid explanations

[Andrew Rycroft]

Hi,

I am looking for an explanation on how sparing is handled by the EVA ( i.e. hot spare disk drives), and an explanation of the VRaid RAID levels and how they are implemented.

With regards
Andrew

[florence mathon lermusi]

hi,

you can have a look on http://h200006.www2.hp.com/bizsupport/TechSupport/DocumentIndex.jsp?locale=en_US&contentType=SupportManual&docIndexId=179166&prodTypeId=12169&prodSeriesId=321347

some info in "best practise" white paper. i attach a white paper on the virtualisation it is a quite old COMPAQ doc...

[Fernando Bandeira]

Hi Andrew,

This is going to be quite lengthy, but I will try to answer as best as possible.

VRAID and sparing are actually quite similiar in the EVA in that they are both implemented at a block level rather than in the traditional manner, which was at a disk level. This means that we now dedicate a certain number of blocks in the disk group for spare capacity and/or for VDISK capacity. So, if you would like the functionality of one spare drive for your disk group, the EVA will reserve an amount of space in the disk group that will allow it to rebuild the biggest disk in the disk group. Both Vdisks and sparing in the disk group are distributed - which means that each block is written non-contiguously across every single drive in the disk group. So in the case of sparing, it makes for very fast rebuilds, because the data is actually being rebuilt on a small portion of all the drives in the disk group. Vdisks work on the same principle. VRAID0 is the method of striping at a block level across every disk in the disk group until the usuaable capacity requested has been reached. The left over space in the disk group can then be used for another Vdisk. VRAID5 is the method of striping with parity in the pool non-contiguously across every disk in the disk group until the usuable capacity requested is reached. VRAID1 is actually more like RAID0+1 in that it stripe-mirrors blocks non-contiguously across every disk in the disk group. All of these VRAID levels can be used in the same disk group at the same time, as we are not dedicating a disk to VRAID5 or VRAID1. Just remember though that the EVA asks for usable capacity when a Vdisk is created, so a 500GB Vdisk will take up approximately 500GB in the disk group, while the same capacity Vdisk but in VRAID5 will take up about 600GB and with VRAID1 it will take up more or less 1TB.

Hope this helps.

Regards
Fernando

[Liam Caffrey]

Hi,

How does VRAID work out when considering the I/O activity that normal RAID incurs?
e.g.
RAID 1 & RAID 10 generates 2 I/Os for each logical write
RAID 5 generates 4 I/Os for each logical write
RAID 5 ADG generates 6 I/Os for each logical write.
Does the same apply to VRAID?

In the context of a SQLServer installation (or any database for that matter), does it make sense to carve up disk volumes for the database transaction logs. Or in other words how can I configure the VRAID (partitions?) on the SAN to ensure that the transaction logs still utilise track-to-track (sequential) seeking rather than random seeking. Does VRAID make all these considerations redundant?


BTW, Fernando's reply is the only explanation of VRAID that I have managed to find anywhere on the Web!

Thanks

Liam

[Leif Halvarsson_2]

Hi,

Liam,
Vraid is rather different from ordinary RAID and "rules" from ordinary RAID systems does not necessary apply to Vraid.

For example, Vraid1 does not give better performance then Vraid5. The main advantage with Vraid1 is better security. Vraid1 uses redundant disk pairs (2 disks), Vraid5 redundant storage sets (RSS,6-11 disks). A Vraid1 volume will loose its information if both disks in the same pair fails, a Vraid5 volume if 2 disks in the same RSS fails.

[Uwe Zessin]

RSS (redundant storage set) is a characteristic at the disk group level, not of individual virtual disks. A whole physical disk drive is a member of a single RSS.

[Liam Caffrey]

Where can I read some information on this subject?

What about the I/Os multiples that VRAID5 uses. Does it consume 4 I/Os for each logical write.
Could I mix a RAID10 with a VRAID5 on a SAN? I want to achieve max performance on the database logs.

I don't really follow why VRAID1 gives better security - maybe you mean reliability? Even then I'm not sure I follow.

Regards

Liam

[Bill Costigan]

The same general rules apply for I/O counts for raid and vraid. And yes you can mix raid 0/1 and 5 on the same SAN and in some cases (e.g. the eva) on the same physical disks.

One of the main differences is that in a typical raid implementation the raid configuration is on a disk basis. 4 data disks and a parity disk. With vraid the implementation is based on disk blocks 4 data blocks (written to different disks) and a parity block (written to yet another disk) The difference is that the parity blocks are not always on the same disk. Every disk in the group will have parity and data blocks. And your file isn't limited to a single group of 5 disks but could be spread over a 100 disks with the controller keeping track of where all the data a parity blocks are. From a performance stand point you can get the performance of a 100 disks instead of 5.

However each write still needs to write the data and parity to different disks and perhaps read the other 3 disks in order to calculate the parity block. (or two writes for raid 0/1)

A real question is how smart and efficient is the controller code. Can it reduce I/Os by batching things and using cache. For example if you were sending 4 sequencial blocks the controller could save each block, calculate a single parity block and write the 4 data blocks and the parity block with 5 I/Os total (4 user I/Os = 5 actual I/Os) Or it could write the parity and data block for each I/O (4 user = 8 actual I/Os.) lastly it could do 5 writes for each user write. (4 user = 20 I/Os)

[Liam Caffrey]

Does that mean then that it doesn't really matter how many I/Os happen for each logical write, because VRAID is so efficient at spreading data into blocks across dozens of disks that the spread more than compensates for any "inefficiency" in the I/O multiplier effect of VRAID (i.e. 4x, 6x whatever).

Is there any point in carving out a VRAID1 if all it offers over VRAID5 is better security as Leif suggests.

Can you point me towards some reading material on this. I feel like the guy doing the maze in the dark! Specifically guidelines for carving up disk space from a DBA perspective. That would be really good and would be well worth a 10/10 rating if I could only find the "drop down points menu beside each reply" which doesn't seem to be on my browser instance.

[Bill Costigan]

It does matter how many I/Os you do, and raid 0/1 can acheive better perfomance numbers than raid 5.

Let's assume that raid 5 does take 5 I/Os for each user I/O and raid 0/1 takes 2.

If we have a LUN spread over 100 disks (each of which can do 100 IO/sec.) the Array could do 100*100 I/Os or 10,000 I/Os per second.

Using raid 0/1 (2:1 reduction) the user could do 5,000 I/O per second. Using raid 5 (5:1 reduction) the user could only do 2,000 I/Os per second. Which is of course still 20 times the I/O of a single disk.

But you need to remember that all the systems connected to the EVA are competing for these 10,000 I/Os per second.

Here is a link to a white paper that talks a little about this.

http://h200006.www2.hp.com/bc/docs/support/SupportManual/lpg29448/lpg29448.pdf

The details on how this all works are spread across a number of documents. There seems to be better technical discussions on array perfomance for the va74xx than for the eva. From a concepts point of view, the va and the eva does the same thing. There is of course differences at the detail level.

Both product lines support read and write caching, raid 0/1 and raid type 5 (the va uses double parity)and LUNs spread across a large number of spindles to increase performance. So you can get a basic understanding of the technologies and benefits from documents covering either product line.

[Leif Halvarsson_2]

Hi,

I actually made a number of benchmarks with an EVA 3000, most of them backup/restores but also IOPS (postmark). And I could not notice any better performance with Vraid1.

Unfourtunatly I don't saved the results.

[Uwe Zessin]

It's not surprising. With backup you have large sequential I/Os. If you write a container file to the EVA:

- with VRAID-1 you have 4 data segments that need to be mirrored - that makes 8 write I/Os.

- with VRAID-5 you have 4 data segments and 1 parity segment. In case of a large transfer, the storage array collects a whole data chunk in writeback cache memory and then calculates the parity segment once - no 'RAID-5 penalty'. Now it can send out the writes: 4x data + 1x parity = 5 write I/Os total.