[prev in list] [next in list] [prev in thread] [next in thread] 

List:       ngw
Subject:    Re: [ngw] GW651 RAID 5 or RAID 10 perfomance on SAN
From:       Randy Grein <randygrein () comcast ! net>
Date:       2005-07-30 3:07:36
Message-ID: EDF97431-6EA4-40A0-B7E6-97033A233A95 () comcast ! net
[Download RAW message or body]

Level 1,0 is no longer considered to be a 'complex' arrangement; it  
was one of the configurations I was originally talking about. (You  
refer to it as 1+0; both are correct but it is generally possible to  
configure as a mirrored strip set - bad - or a striped mirror set,  
good.) You are correct regarding the number of spindles being the  
overriding factor but incorrect in the assumption that mirrored  
drives are capable of alternating access. That is entirely dependent  
on the controller algorithm and there are well-known cases where  
simultaneous access is designed in - something mentioned in your  
second reference, although it does utter the mistaken pronouncement  
that RAID 5 has "Highest Read data transaction rate". I wouldn't put  
too much faith in that particular reference; while much of the  
information is correct some is rather out of date. Odd, considering  
the section on SATA RAID... . You might check out the original RAID  
paper at Berkley; there is a second one titled RAID II that I used in  
an article on the subject about 8 years ago. You can find the RAID II  
paper here: http://www.cse.ucsc.edu/~elm/Papers/vlsidesign93.pdf .  
Here's a pointer to some earlier work; note that it calls out a  
different RAID 6 than uses that name today and references a  
questionable RAID 7 level.

You need to correct your second paragraph, BTW. The SCSI command set  
allows for a number of performance-enhancing features such as command- 
tag queuing. In other words, multiple requests can be made, the drive  
will resort the order as necessary and service the requests in the  
most efficient fashion. This is essentially the same concept as  
Novell's elevator seeking. Command tag queuing is just starting to  
come into play in SATA drives but in general it can be assumed that  
the RAID controller will handle reordering and multiplexing requests.  
If an array had to wait for last block retrieval before requesting  
the next block performance would be poor indeed.

Randy Grein, Master CNE, CCNA

On Jul 29, 2005, at 6:28 PM, Eric Rothweiler wrote:

> In short, I was pointing out that mirrors only give a choice of 2
> spindles to read from.  Stripes give more than 2 so a third spindle
> can be fetching block 3 while blocks 1 & 2 are being read from
> spindles 1 & 2 in whatever order the controller is smart enough to get
> the disk to provide the data.  So to be fair I need to retract raid 0
> from that statement but since 0 offers no data redundancy most shops
> do not consider raid 0 by itself.
>
> The data bus is faster than the drives ability to position the read
> head where it's needed.  A simple way to look at it is that you have 8
> pieces of data to read.  In most cases the 8 blocks will not be side
> by side so the drive will need to reposition the heads to get to each
> block.  in a 2 spindle situation the best you will have is a read of
> alternating blocks between physical drives.  In a full raid 5 set (or
> 1+0 which is what I assume you mean by 1/0) the controller will
> instruct each spindle to fetch a different block.  Once it gets block
> 1 from the first it will tell it the next block it needs (block f) and
> while that spindle moves around it will get the next block from the
> next spindle so that block f is in buffer when the controller gets
> back to it.
>
> Reference:
> http://www.everything2.com/index.pl?node=RAID
> http://cahtech.beigetower.org/howto/raid.php
> http://www.gen-x-pc.com/more_raid.htm
> http://www.prepressure.com/techno/raid.htm
>
> notice that each article credits raid 5 as being fastest (barring a
> complex arrangment that opens the possibility of raid 1+0 or 5+0)
>
> On 7/29/05, Jeffrey Sessler <jeff@scrippscollege.edu> wrote:
>
>> I've never seen any performance papers that indicate that RAID 5  
>> has better read performance vs RAID 1/0. I checked my EMC  
>> performance white paper and it indicates that read performance is  
>> equal between the two given the same number of disks.
>>
>> best,
>> jeff
>>
>> Jeffrey D Sessler
>> Assistant Director of Technical Services
>> Scripps College
>>
>>
>>>>> jetadmin@gmail.com 07/29/05 5:23 AM >>>
>>>>>
>> Not quite accurate.  Raid 5 will give better read performance in most
>> cases because each spindle can be seeking the different sectors
>> simultaneously.  In Raid 0 or 1 each seek in done is serial  
>> instead of
>> parallel.
>>
>> On 7/29/05, Randy Grein <randygrein@comcast.net> wrote:
>>
>>> One small point is that all RAID
>>> levels will have the same read performance, the difference is write
>>> performance which will vary depending on the RAID level and write
>>> size. If anyone is interested in reviewing, drop me a line.
>>>
>>
>>
>


[prev in list] [next in list] [prev in thread] [next in thread]