Which is faster? 4x10k SAS Drives in RAID 10 or 3x15k SAS Drives in RAID 5?

Question

I am reviewing quote for a server upgrade. (RHEL).

The server will have both Apache and MySQL on it, but the reason for upgrade is to increase DB performance. CPU has been upgraded massively, but I know that disk speed is also a factor.

So RAID 10 is faster performance than RAID 5, but how much difference does the drive speed make?

(The 15k discs in the RAID 5 config is at the top of my budget btw, hence not considdering 4x15k discs in RAID 10, which I assume would be the optimum.)

Solution 1

Here's the corner-of-the-napkin calculation:

10K drive - 150IOPS, 15K drive - 180IOPS (these numbers are used for example by EMC for array sizing).

RAID10 read performance - 600IOPS. RAID5 read performance - 540IOPS (no penalties, straightforward multiple).

RAID10 write performance - 300IOPS (each block goes to two spindles). RAID5 write performance - 135 IOPS (you have 540IOPS potential with three disks, but each write actually consists of two reads (data and parity) and two writes (new data and new parity)).

These numbers, of course don't take into account controller caching and optimizations, so RAID5 will probably be slightly better than 1/3 of RAID10 performance on write. But this is a good ballpark number anyway.

Of course, if this is a data warehouse, you won't care about write performance. Otherwise, definitely go with RAID10.

Solution 2

The advantage of RAID 10 will be in write performance, so the correct answer really depends on whether or not your database is doing a lot of writing to disk, and whether or not performance matters with this writing.

Bear in mind that RAID 10 can offer a layer of resilience that RAID 5 doesn't have; you can afford to lose 2 disks (provided they're not a mirrored pair) due to hardware failure with RAID 10, but that will kill your array with RAID 5. This may outweigh any read performance gains from 15k RAID 5, and may also outweigh the disadvantage of having less usable storage.

Solution 3

Currently my answer would be none of them. Unless you need lots of bulk storage go for SSD. For database applications they beat the spinning disks easily, raid or no raid. The cost per gig is just a little bit higher than the 15k disks today, and you don't need any raid solution unless you feel that they don't offer enough space.

If that is not an alternative my answer would be a firm "it depends". The performance will depend a huge lot on your raid controller, probably more than your hard disks.

---

edit: Compare to the IOPS calculations in other posts here:

Intel X25-M G2 180G:

• Up to 35K IOPS Random read (4k)
• Up to 8.6K IOPS Random write (4k)
• Up to 250MB/s Sequential reads
• Up to 70MB/s Sequential write

You won't ever get close to those speeds for random load (which is what most DB load is) on a spinning disk.

Comments

[Maximus Minimus]

It's a pity that you can't do 4 x 15k disks in RAID 10 because performance is really sweet.

[TomTom]

A pity mySQL seems to have so bad analysis tools - using SQL Serve you could pinpoint your bottlenecks with exact numbers.

[Helvick]

There is no parity disk in RAID5, the parity is striped and distributed so all spindles can be used concurrently for reads. RAID with a dedicated parity disk is either RAID 3 or 4 and they are rare for this and other reasons. The RAID 5 IOPS nos. Should be 540 (read) and 135 (write) assuming the RAID 5 write penalty is 4, which is generally true for random small IO when cache is saturated.

[TomTom]

Seconded. SD, RAID 5 - hands down beats the hell out of anything you do with physical discs.... AND you get bigger capacity at the same side.

[Sergei]

"The cost per gig is just a little bit higher than the 15k disks today" - can you show where do you get this data from?

[John Gardeniers]

-1 for "you don't need any raid". SSDs fail, just like real drives. Perhaps you've forgotten the principle reason to use RAID.

[pehrs]

Eh? I really really really hope you are not claiming that RAID is there for backup purposes. RAID is not backup. Raid MIGHT improve uptime, but it is also a single point of failure. I have had more data loss from failed RAID controllers than from failed drives during my career.

[pehrs]

Looking through the price lists we use for ordering. SAS disks at 15k at 60-80 gb gives prices of around 2-5 USD/GiG. Similar sized MLC SSD disks costs 3-6 USD/GiG. Intel Intel X25-M G2 are around 4 USD/GiG.

[John Gardeniers]

@pehrs, I didn't even hint at RAID being used for backup. What I was alluding to and what the "R" in RAID stands for is redundancy, which you've discounted as not being required.

[pehrs]

@John Is it required? Should anybody be encouraged to trust RAID redundancy? In my experience raid is fine for improving up time and performance on low quality (customer) spinning disks. My experience is also that SSD disks perform well enough that the raid controller becomes a bottle neck. Finally my experience is that raid systems have nasty failure modes and should be removed whenever possible.

[JESUS ESPINOSA]

Thanks, RAID 10 sounds like the one to go for, it's a 10% cheaper option anyway. I'm hoping the CPU (Dual Quad Core Intel Xeon 5420) will make a huge difference as will being on a 64 platform (so I can take advanatage of increased useable RAM). The only niggle is that our database is heavily SELECT (read) statments rather than writes.

[Maximus Minimus]

@pehrs - have you ever had a single disk fail in a server? Thought not.