Raid 0: Is this as good as it gets for 2 drives?

[mechBgon]

Originally posted by: pm
To me, if a decent backup solution is implemented, then availability of the system becauses important enough that RAID 0 still looks not great.

In one application that I am very familiar with here at work, we had a cluster of 150 PC's each with two drives. The first drive had the OS and application files on it. The second drive was used as a temp disk for the application. If either drive died, the node was out of commission. In this cluster, we saw a failure rate per drive of about 8% over a year - about 25 drives out of 300 died in the first year. These drives were used 24/7 and were thrashed pretty hard with data constantly being read and written to them. They were all 10krpm SCSI. The number of failures got better for the years that followed but we had 8% fail within the first 12 months - most failed within the first 4 months. Data integrity wasn't a problem - the cluster was running applications using data pulled from servers over the network so all that was lost was the job being run. But since we had an IT guy fixing these computers in his spare time while doing his regular job, in general we had about 10% of the cluster idle/dead waiting for an HD swap, reformat and OS build.

Whether or not RAID 0 is "scary" is a matter of perspective. If you have a couple of computers at home and you have a good backup solution then, "scary" is definitely an over-exaggertion. But if you have a cluster of computers and you want as much uptime as possible, then scary is not too much of an exaggeration - particularly if the reliability of the cluster will impact the end of the project. When I finished with that project, one of my "learnings" that I put in my post-project report was to try to get everything on one hard disk instead of two to minimize downtime.

pm, what drive-cooling provisions, if any, were used on those systems? Two active 10k SCSI drives stacked in proximity to eachother can get fairly hot without forced-air cooling.

 

[AgaBoogaBoo]

Never fear, Aga is here!

I'd reccomend getting 2 15,000 RPM drives (Fujitsu MAS, review at www.storagereview.com) and setting those up on Raid 1. One other option is Raid 1 + 0 with 2 drives. What you'll do is partition each drive into 2 halves. This way you have 4 partitions total. Now take one partititon from each and the other to make a Raid 1 + 0 setup. Heres the performance gain: When only reading files for the most part, you will get your Raid 0 performance but when you write files you will also get Raid 0 performance. The only catch is that once you read and write equally then you will ge the speed same as a single 15,000 RPM drive which in itself is faster than most 7200RPM IDE drive Raid 0's. This is all theoretical and I have never worked on a Raid setup myself so I can not say for sure but if the program supports Raid with partitions this will work.

 

[Noriaki]

If you use RAID0 I suggest you keep the drives well defragged if you want to see an increase in performance.

RAID0 will not increase drive seek performance. It increases STR by quite a lot no doubt about that, but often it's the seek performance that will dictate your overall performance, not STR, so even though you have huge gains in STR you won't see much gain in performance.

Unfortunately a lot of hard drive benchmarks put alot more weight on STR than most regular usage will, so RAID0 looks more attractive from most benchmarks than it really is.

That coupled with everything that Patrick already raised about reliability concerns would make me think long and hard about RAID0.

It is possible you have some applications that will be heavily STR limited in which case you will see big gains by a RAID0, but usually the STR increases just won't help you that much.

If you want to increase seek performance, go for a high speed SCSI drive, 10k RPM or 15k RPM will give you much better seek performance than a pair of striped ATA drives, which will win on STR depends on the drives Areal density, but if all the drives are new likely the ATA RAID will have some edge in STR, though likely not enough to offset the Seek Performance gain from SCSI.

All in all I'd take a single SCSI drive over an ATA RAID for Workstation performance in most cases, but it is somewhat application domain dependant.

 

[FishTankX]

Samsonoid, one thing I would like to point out is that it's possible for a WD raptor to perform better than two 7200RPM SATA harddrive's in RAID0. It all depends on what work your doing.

That being said, the raptor is a very sweet drive from a performance perspective. I'd say a single raptor would buy you sweet performance.

 

[Samsonid]

Originally posted by: AgaBooga
Never fear, Aga is here!

I'd reccomend getting 2 15,000 RPM drives (Fujitsu MAS, review at www.storagereview.com) and setting those up on Raid 1. One other option is Raid 1 + 0 with 2 drives. What you'll do is partition each drive into 2 halves. This way you have 4 partitions total. Now take one partititon from each and the other to make a Raid 1 + 0 setup. Heres the performance gain: When only reading files for the most part, you will get your Raid 0 performance but when you write files you will also get Raid 0 performance. The only catch is that once you read and write equally then you will ge the speed same as a single 15,000 RPM drive which in itself is faster than most 7200RPM IDE drive Raid 0's. This is all theoretical and I have never worked on a Raid setup myself so I can not say for sure but if the program supports Raid with partitions this will work.

Using partitions .... mmmmmm... interesting !
Tell me some more.

If we cheat using said partitions for a 1/0 array do we still get the *full* benefit of preserving the data if and when one of the two drives fail? (is there a catch in this area) Can the system still run with one drive only ?

Also, do I understand this correctly ? If we have two 40Gb drive (a total 80Gb) and put them into a Raid1/0 with partitions then we only have 20Gb available for use? That means we are using only 25% of the total 80Gb available capacity?

(That is probably the same as using 4X40Gig drives in Raid1/0... we only get to use 40Gb out of 160... which is still 25%)

Is my math correct?

In a Raid1 we get to use 50% of the total capacity.

So comperatively speaking Raid1/0 costs twice as much as Raid1

 

[pm]

Originally posted by: mechBgon
B]pm[/b], what drive-cooling provisions, if any, were used on those systems? Two active 10k SCSI drives stacked in proximity to eachother can get fairly hot without forced-air cooling.

They were all the same model IBM server - I can't recall the model number but if you are especially curious I can look it up on Monday. Dual Xeon's with 2GB of RDRAM on i850 boards. Cooling was the stock setup from IBM. We didn't mess with them at all after we unpacked them. It, had, as I recall, a thermally controlled massive fan on the front of the case and a fair amount of ducting. The drives were IBM's and they didn't seem to get especially hot... more than warm, but not what I consider hot.

 

[Samsonid]

There is a bit speculation floating around of which combination of Raid vs Single SCSI vs single Raptor vs Raid SCSI is the best price/speed performer.

Is there a link to a site that shows actual real life performance tests in direct comparison between ATA-Raid and SCSI combinations.

I used to have a link but can't find in any longer.


Thanks

 

[Noriaki]

The size of a RAID0 is the sum of the sizes of the disks.

Half your data is stored on one disk, half on the other (in the two disk case), and data is read or written by alternately stripping off each.

RAID1 makes duplicates of the data on all the disks in the set, so the size of the RAID1 is the size of one disk.

RAID1/0 with four disks in a 2x2 set would give you 50% of the sum of the disk sizes.

Suppose disk1 and disk2 are a RAID0 set, so you get (disk1+disk2) total size.

Then disk3 and disk4 are another RAID0 set, same size.

Then you mirror these two sets just pretend each of those RAID0 sets is 1 "virtual" disk so if you mirror those you will get data duplicated in each of the two RAID0 sets, resulting in RAID0+1 or RAID10.

RAID0 = 100% disk space usage.
RAID1 = 50% disk space usage (in the 2 disk case).
RAID0+1 = 50% disk space usage (in the 2 RAID0 set case).

 

[Sideswipe001]

For what it's worth-

I've run IDE (or SATA) RAID 0 for over 2 years now on my home system, without a single loss of data or drive going bad.
One of my good friends is at about a year and a half without any issues as well.

That's how "scary" running RAID 0 is for me.

 

[Noriaki]

Originally posted by: Sideswipe001
For what it's worth-

I've run IDE (or SATA) RAID 0 for over 2 years now on my home system, without a single loss of data or drive going bad.
One of my good friends is at about a year and a half without any issues as well.

That's how "scary" running RAID 0 is for me.

Don't take this the wrong way, I'm not meaning to bash you or anything, but anecdotal evidence in this sort of situation is meaningless.

You can *ALWAYS* find someone who has had no problems whatsoever with their RAID0 set and someone who has had no end of problems.

The question is how statistically significant is it? The statistics are that running RAID0 is more likely to result in data loss than running a single disk.

Though there are always ways to protect yourself against dataloss if you want to run RAID0 and that is important to you. But then there is downtime... It's all a matter of what your priorities are.

Anyhow...the real reason I posted again was to point out http://www.storagereview.com/ I haven't been to the site in a while, but as memory serves they have pretty good reviews of various interfaces and drives.

 

[Samsonid]

Originally posted by: pm
To me, if a decent backup solution is implemented, then availability of the system becauses important enough that RAID 0 still looks not great.

......

Whether or not RAID 0 is "scary" is a matter of perspective. If you have a couple of computers at home and you have a good backup solution then, "scary" is definitely an over-exaggertion. But if you have a cluster of computers and you want as much uptime as possible, then scary is not too much of an exaggeration - particularly if the reliability of the cluster will impact the end of the project. When I finished with that project, one of my "learnings" that I put in my post-project report was to try to get everything on one hard disk instead of two to minimize downtime.

Downtime is quite important. In a Raid0 if one of the two drives is a "lemon" and fails within two months then the whole system is down for the full length of time needed to acquire a replacement drive. At that point both drives have to be reformated and started again from scratch and use the back-up solution to put the data back in.

(In raid0 the life of the system is as much as the weekest drive and we have to take into account the productivity loss from the inoperable station and the cost/time to repair it).

If two Raptors in Raid1 read/seek (random) twice as fast as a single Raptor then it is good news.... but somehow I think there is a catch... it is probably just 1.2 times faster seek (instead of 2X).

Does anyone have the facts on read/seek times on a Raid1 versus single SATA or SCSI ?

 

[Samsonid]

Originally posted by: Noriaki
The size of a RAID0 is the sum of the sizes of the disks.

Half your data is stored on one disk, half on the other (in the two disk case), and data is read or written by alternately stripping off each.

RAID1 makes duplicates of the data on all the disks in the set, so the size of the RAID1 is the size of one disk.

RAID1/0 with four disks in a 2x2 set would give you 50% of the sum of the disk sizes.

Suppose disk1 and disk2 are a RAID0 set, so you get (disk1+disk2) total size.

Then disk3 and disk4 are another RAID0 set, same size.

Then you mirror these two sets just pretend each of those RAID0 sets is 1 "virtual" disk so if you mirror those you will get data duplicated in each of the two RAID0 sets, resulting in RAID0+1 or RAID10.

RAID0 = 100% disk space usage.
RAID1 = 50% disk space usage (in the 2 disk case).
RAID0+1 = 50% disk space usage (in the 2 RAID0 set case).

Thank you for this. I had it mixed up.

 

[Samsonid]

Originally posted by: AgaBooga
Never fear, Aga is here!

One other option is Raid 1 + 0 with 2 drives. What you'll do is partition each drive into 2 halves. This way you have 4 partitions total. Now take one partititon from each and the other to make a Raid 1 + 0 setup. Heres the performance gain: When only reading files for the most part, you will get your Raid 0 performance but when you write files you will also get Raid 0 performance. The only catch is that once you read and write equally then you will ge the speed same as a single 15,000 RPM drive which in itself is faster than most 7200RPM IDE drive Raid 0's. This is all theoretical and I have never worked on a Raid setup myself so I can not say for sure but if the program supports Raid with partitions this will work.

I don't think this is going to work. It sounds exciting but it is not.
Performance on "write" operations will be slower than a single drive system.

The "1" component of the Raid means that it is going to mirror the full packet of data to each drive.
Now...for the "0" component, it also has the additional job of splitting each mirrored packet in two (this is like "deliberate fragmentation" )

Reading is going to be slow too, because now for each packet it has to access the same disk in two different sections to recompile the fragmented data (half from one drive and half from the other).

This solution will be *slower* than Raid 1 alone.


Very doubtfull that this is a viable solution. It it was exciting while it lasted though.

 

[mechBgon]

Why not build two prototype systems, and pick up two WD360GD's, one Cheetah 15k.3 (or Atlas 15k, or Fujitsu MAS-series), one LSI Logic U160 controller and a SCSI cable/terminator? You can test the different possibilities (single WD, RAID0 WD, single SCSI) and see what feels or benchmarks fastest for what you actually do with these systems. Hypermicro carries all that stuff. Worst-case scenario, you have one or two drives and maybe a controller left over.

Sight-unseen, I would go for SCSI, preferably 15000rpm if the higher cost per Gb isn't a problem for you. The Cheetah 15k.3 has fluid bearings and can be expected to idle quieter, if that's a factor at all.

 

[charlie21]

Originally posted by: AgaBooga
Never fear, Aga is here!

I'd reccomend getting 2 15,000 RPM drives (Fujitsu MAS, review at www.storagereview.com) and setting those up on Raid 1. One other option is Raid 1 + 0 with 2 drives. What you'll do is partition each drive into 2 halves. This way you have 4 partitions total. Now take one partititon from each and the other to make a Raid 1 + 0 setup. Heres the performance gain: When only reading files for the most part, you will get your Raid 0 performance but when you write files you will also get Raid 0 performance. The only catch is that once you read and write equally then you will ge the speed same as a single 15,000 RPM drive which in itself is faster than most 7200RPM IDE drive Raid 0's. This is all theoretical and I have never worked on a Raid setup myself so I can not say for sure but if the program supports Raid with partitions this will work.

One big problem with this. Regardless of how many partitions you have, you have to remember you still only have 2 physical drives. You would lose the redundancy of a true 0+1 setup. If one drive fails, your array is still fvscked.

 

[FishTankX]

Originally posted by: Samsonid

Originally posted by: pm
To me, if a decent backup solution is implemented, then availability of the system becauses important enough that RAID 0 still looks not great.

......

Whether or not RAID 0 is "scary" is a matter of perspective. If you have a couple of computers at home and you have a good backup solution then, "scary" is definitely an over-exaggertion. But if you have a cluster of computers and you want as much uptime as possible, then scary is not too much of an exaggeration - particularly if the reliability of the cluster will impact the end of the project. When I finished with that project, one of my "learnings" that I put in my post-project report was to try to get everything on one hard disk instead of two to minimize downtime.

Downtime is quite important. In a Raid0 if one of the two drives is a "lemon" and fails within two months then the whole system is down for the full length of time needed to acquire a replacement drive. At that point both drives have to be reformated and started again from scratch and use the back-up solution to put the data back in.

(In raid0 the life of the system is as much as the weekest drive and we have to take into account the productivity loss from the inoperable station and the cost/time to repair it).

If two Raptors in Raid1 read/seek (random) twice as fast as a single Raptor then it is good news.... but somehow I think there is a catch... it is probably just 1.2 times faster seek (instead of 2X).

Does anyone have the facts on read/seek times on a Raid1 versus single SATA or SCSI ?

No form of RAID ever has the power to affect seek time. Period.

The only thing affected in transfer rate.

RAID1 has identical read performance with RAID0 but half the write speed.