RAID10 (16 x 256GB SSDs) in PowerEdge R720 for Build Server

Question

I'm trying to put together specs for a top of the line Dell R720 (running Windows Server 2008R2) build server (C#), and I'm unsure what the best hard drive / RAID setup would be.

Our space requirements are:

• C (OS) - 100GBs
• E (SOURCE) - 600GBs
• F (OUTPUT) - 850GBs

Utilizing the Dell R720, I was going to go for:

• 16 x 2.5" drive bay
• 16 x 256 Crucial m4 SSDs
• H710P HW Raid Controller
• 2 x RAID1 (C drive)
• 6 x RAID10 (E drive)
• 8 x RAID10 (F drive)

Without investing in ioDrives, does the setup above appear ideal for the fastest write speeds possible? I understand the risks we're taking with non-OEM, non-enterprise SSDs, but my goal here isn't stability, it's to attain the best write performance possible for F: drive.

---EDIT---

Some points @ewwwhite raised:

Do you know what your realistic I/O needs are? ##

Unfortunately I do not, other than:

1. Copy data from network to E: drive as fast as possible (10GbE, 6Gbps SATAIII)
2. Read data from E: drive into Visual Studio.NET (2008) Compiler
3. Write data sequentially (as 4GB+ binaries) to F: drive

Mostly random read/writes? Sequential reads/writes?

I'm not familiar enough with how linear our C# build process reads / writes, but I believe it is more sequential than random.

How much room and storage space do you need

100GBs for the OS, 600GBs available for SOURCE, 850GBs available for output. About 85% utilization for the SOURCE and OUTPUT volumes.

Answer 1

Most RAID controllers have a cap on how many solid state drives can be accommodated at a time. There is a point of diminishing returns. For HP Smart Array P410 controllers, it's about 6 x SSD's in RAID0.

Do you know what your realistic I/O needs are? What is the working set of data and the expected data output size? What would the I/O pattern be? Mostly random read/writes? Sequential reads/writes? How much room and storage space do you need? This matters a bit in the design.

I wouldn't eschew enterprise drives at this point just yet. There are differences between various SSD offerings, and some devices are capable of far greater throughput than others. For example, a single ZeusRAM SSD drive has 8GB capacity, but can sustain 100,000 IOPS at 800+MB/sec and microsecond latencies and won't wear-out. Or reconsider a FusionIO drive. It's on par with the price of the solution you're proposing and would be a far more efficient approach.

Can you help us understand what you realistically need?

Answer 2

I think this sounds like a good storage setup for a build server. It seems very well balanced for the needs of a typical build server. A RAID-0 setup or FusionIO drives could be faster, but will provide less redundancy. I wouldn't use RAID-0 for a serious development shop, as your build-server would then be unavailable in case of a single disk failure. Similarly, a single FusionIO card constitutes a single point of Failure, that you must weigh against how critical, the server is for your developers. FusionIO cards are quite stable, and have a significant degree of internal redundancy, but I have seen them fail.

Answer 3

I don't know if this should be a comment or an answer, but I happen to have a stack of somewhat similar machines configured as follows:

• Dell R720's with two E5-2643's and 64GB of RAM (eight 8GB sticks.)
• Dell PERC H710p Mini (a rebranded LSI MegaRAID SAS 2208)
• Seven Dell MZ5EA100HMDR SSDs. (rebranded Samsung SM825s)

The seven SSDs are configured as a RAID5 with one hotspare. These machines have been a running weird, custom database software pretty much 24x7 for the past three years:

root@stagingtest0:~# hdparm -t /dev/sda

/dev/sda:
      Timing buffered disk reads: 4204 MB in  3.00 seconds = 1401.17 MB/sec

I don't know any IOPs measurements off the top of my head, so unless anyone suggests one, the hdparm benchmark will have to do.

smartmontools can interrogate the drives correctly using -d megaraid,X and we track the SSD wear. Still no failures over 3 years and ~90 SSDs. If I'm reading the SMART attributes correctly, though, some SSDs are showing 60-65% worn out. smartmontools for windows can't interrogate LSI controllers to the best of my knowledge, so you may be out of luck there.

Not knowing more specific details, I doubt RAID10 will do you much benefit. If the data is important, however, I'd recommend running hotspare(s). In our testing, the SM825's resync'd to the hotspare in 15 to 30 minutes. With RAID5 H710P was more than fast enough for our needs.

We also found the enterprise DRAC very worthwhile.

Oh, and a coworker woefully overestimated the power supplies. We ended up with dual 750's when we could've easily used dual 350's.

Good Luck!