Raid 0.5 A Great Idea

[americanbrian]

OK, so I'm sitting at work reading toms hoping that there is nothing promoting dell today, when BAM 8O an idea comes to me as I read these cool new external drive functions.

Everyone knows that raid 0 is great for gamers that care only about performance (load times) and nothing about safety. Conversely you have eggheads who do relentless work on computers and would commit suicide if they found that there new ... was corrupted.

I know something roughly similar has just been released to the market but this idea is going to fill in the gaps.

Your controller first stripes your array in the same fashion as RAID 0, writing/reading to/from alternate stripes on either drive. This is how the array manages to pull out some extra performance.

In my proposed implementation the stripes are spaced with an extra one in the middle which remains blank for the time being while the drives resources are being utilised for I/O with the user.

During idle time the controller then mirrors the stripes to the empty ones on the opposed drive in the array.

True data loss could still happen but not on nearly as big a scale as the catastrophe of a raid 0 all data gone forever kinda way.

It is easier to represent in a picture, which I would be willing to do but not until later, especially as I just won't bother if this is made a mockery of.

 

[MoNeY3865]

What you are basically describing is a RAID 1 array modified to delay some of the writes.

Good RAID controllers will get read times from a RAID 1 array that are almost as fast as read times from RAID 0 arrays because they will read different stripes of data from each identical drive. RAID 0 arrays are always faster when writing though because you only have to write the data once.

Your idea for copying the data during downtime is neat, but there is still a slight risk that the drive could fail before there was downtime. This is because I assume the "downtime" would be something like a nighttime chore, otherwise why not just use RAID 1 with a good controller.

I think the simper solution would be to just use a RAID 0 array for speed and then just backup the data to another drive during the downtime.

 

[dean7]

Not to mention the fact that he was forgetting about RAID 5, which offers performance + redundancy.

And yes, I have a RAID controller running RAID 1 and the read performance is amazing.

 

[eRazor]

RAID-5 performance suffers badly in a scenario involving random writes. The sequential read performance of RAID-5 is only marginally better than a single disk when used in the minimal 3 disk configuration.

RAID-0 (striping) with regular backups is the best performer. If you can afford it, RAID-0+1 (mirrored striped set) is the best overall for performance and fault tolerance.

RAID 5 has it's place, but IMHO, it's not a workstation solution if you're looking for performance.

Where I see RAID-5 being used is in remote locations where there is no tehnical people to service the machine should failure occur. In such a case I would use a 3 disk RAID-5 with a pair of hot spares or a RAID-6 (striped with dual distributed parity) with 2 hot spares. That would give me time to get to the remote location without having to worry about data loss and down-time.

 

[dean7]

RAID 5 suffers from random writes? Proof? Some sort of example?

With 4-6 drives RAID 5 performance is great, and with less drives it really depends on how nice your RAID controller is.

 

[SomeJoe7777]

RAID-5 performance suffers badly in a scenario involving random writes. The sequential read performance of RAID-5 is only marginally better than a single disk when used in the minimal 3 disk configuration.

You can't make a blanket generalization of a technology like that. RAID 5 performance is tied to the controller, pure and simple.

The 3Ware 9650 can sustain 600MB/sec RAID 5 writes on an 8-drive array.

One of my 3Ware 9500S controllers can do 170MB/sec RAID 5 reads and writes on a 3-drive array.

RAID 5 speed depends entirely on the controller computing power, design, interface, and cache. Cheapie motherboard controllers have low computing power, consumer-level design, and no cache. Is it surprising that they have low performance? Of course not. You get what you pay for.

 

[americanbrian]

Thank you all to those who commented. I have seen the results and only one of you thinks its a bad idea.

To clarify though, by downtime i mean that a threshold could be set for monitering the I/O utilisation of the drive so that if you are say watching a movie, or generally not taxing the disks too much it uses the overhead capacity to complete these mirror stripes.

If loading massive textures during the game, or ripping from ultrafast media/disk it would simply function with the non-redundant feature. This could be more useful given the new ability to raid 0 external drives. There wouldn't really be much point if you were limited by the 1x speed of the internal drive.

I accept that the level of security is not as high as raid 1, and the write times may not quite reach the performance of the Raid 0, perhaps a comfy medium could be reached.

 

[eRazor]

Yes, in fact I do have some experience with Exchange and SQL (different servers) on both Adaptec 2100S and 3100S controllers.

Now, I don't claim to have exact numbers in front of me as it's been several years since these two machines were put into production. The numbers have long since been forgotten.

That said, loadsim on exchange tested with a 2100S in RAID-5 had a MDPM (Messages delivered per minute) decrease in the order of 25 percent over the same controller in a RAID-0. The numbers changed to ~ 20% in a RAID-1 setup with a hot spare.

Similar results were obtained in another location using a 3100S.

Both of these cards were around the $1,000.00 mark when purchased.

I don't claim to be a RAID expert, but that's what I base my statement on. In that application, RAID-5 isn't the answer.

YMMV.