Moving to SSD storage from Platters: Where does RAID come in?

Question

When moving to SSD, do you keep RAID10?

We have a server with internal SAS disks in a big array:

16x600GB RAID 10

This gives us ~4.6TB of storage. It is fast (this is SQL Server).

Our SQL database is:

-- 600GB MDF -- 200+GB/Day of log files -- Heavily transactional -- Heavy read and write workload

We are looking to move to SSD (for the usual reasons: Speed, speed, and performance). We expect to use Write Intensive 400GB SSDs. (They seem to be the best cost-capacity ratio)

We believe that on SSD, there is less need for RAID10 (a big reason for RAID10 is to boost throughput, but SSDs solve for that natively).

RAID5, with spinning disks, is a performance disaster for SQL (it only made sense when disks were super expensive).

In the Dell world, we would use the "Write Intensive MLC" SSDs.

But for SSD, what is the good way to go?

• RAID10? Rather expensive on SSDs...
• RAID5? Is the write penalty manageable in an SSD environment?

In all cases, we will have one or two hot spare SSDs in the chassis.

What is the appropriate approach?

-- We understand that RAID10 is the cadillac -- We would prefer to reduce the device count (given that SSDs solve the performance issues that RAID10 is used to solve) ---- Reducing device count helps on the $ side, and on the expansion side (use fewer drive bays...)

But, will going to RAID5 on SSDs penalize us in some way?

-- We know that RAID5 performance, on platters, goes to hell when one drive dies. -- And RAID5 has poor write performance: Is that hit coming from (a) the controller (calculating parity), or (b) from the need to wait for both writes (data block and parity block) to complete? If it is from 'b' then the higher IOPS of SSD should solve, right?

Answer 1

Honestly, there are so many potential variables that the best approach IMO is to benchmark different setups and let the results help you decide. Software vs hardware raid, different performance characteristics between vendors.. whether you have a battery backup unit on the raid controller to accelerate writes all factor in.

I suggest you define what your requirements are for the RAID, how much redundancy you're looking for, what you're optimizing for (read speed? write speed? simultaneous read speeds?). Once you have that you exclude the RAID levels that won't meet your needs, then benchmark the remaining ones that can meet your requirements.

Anything less than that is basing decisions on anecdotal evidence and hunches. It'll probably still work but if you want to know what's best for your situation you'll have to figure that out on your own.

Answer 2

If you need the capacity, you should still use the RAID solution that helps you meet your capacity, redundancy and IOPS requirements. Note: you didn't specify the SIZE of the SSDs you wish to use...

In many cases, that means continuing to use RAID 1+0.

I feel like your question is really something like, "make me feel better about using RAID 5"

You can do whatever works. I'd caution against hot-spares with the cost of this type of SSD. Keep a cold-spare handy and be done.

Answer 3

Even though RAID 5 deliver worse write performance than RAID 10, it's very probable your application will have it's performance greatly improved from using mechanical drives when you migrate to Enterprise SSD's.

Having some knowledge of your read/ write mix would be very useful to calculate the difference of performance between RAID 5 vs 10 (is approximately equal to the % of writes). I work at Dell so maybe I'm biased, but I think asking your Dell sales rep to run a DPACK collection would be the easiest way to uncover this information, along many other that you probably will find interesting, like total IOPS, storage and network bandwidth, etc.

The performance difference between RAID 5 and RAID 10 is mainly in write operations. You shouldn't perceive any significative difference in read operations between any RAID level. For sizing calculation of a predominantly random workload (where most OLTP databases fall) you could consider that RAID 5 has about half the write performance of RAID 10 when using same hardware. For more information on the rationale behind sizing of RAID check this: http://www.storagecraft.com/blog/raid-performance

About the SSD version, it depends mainly in how many write operations (write IOPS) you have. Read intensive, mix use and write intensive is a form of classification of SSD's drives according to their write performance. It may pass the impression that a RI can surpass the Read performance of a WI SSD, but this is far from true. You can check this in , comparing S3510 (classified as RI at Dell), S3610 (MU) and S3710 (WI). The price per GB increase when going from RI to WI, and in most of the cases RI offer the best price per performance (except if you got a lot of write operations).

Answer 4

This one is easy, make a raid10 bcache from SSD's configured to write smaller than 4MB to cache. Make another raid10 of mechanical drives and cache it with bcache.

I have done this with IT mode HBA card 24 x 12TB disks and 4x500GB ssd's for a an office which perform heavy rendering operations , 2d and 3d design on this server. It works flawlessly ...

Note that, 32 core dual xeon c600 series chipset with 128 GB ram. I have even virtualised some other OS with KVM within server to let them organize their files and permissions the way they are used to. I have used virtio NET and virtio Disks so that even virtualised ones can have tremendious I/O and throughput. Which easily saturates dual port 802.3ad bonded 10G network card...

I strongly recommend this.

!!!: Do not even think about mdadm raid 5 raid 6 or btrfs they are buggy. And have very long rebuild times. Wrongly configured ssd only raid bypasses ssd SLD cache. You will be looking extremely low write values if you do not adjust it correctly and write a script to fstrim the disks regularly ...