Intel Atom C2750/C2758 powerful enough for RAID 6?

[skoehler]

Hey,

I was going to build a RAID6 system with 4 disks based on a SuperMicro A1SRi-2750F or 2758F board. Obviously, the Atom (even though being octo-core) is not the most powerful CPU in the world. But the question is: can it handle the computations needed for reading/writing a RAID6 with 4 disks?

There are 2 potential bottle necks:
1) the computations for RAID6 exhausting the CPU's computational power
2) the bandwidth between the main memory and the sata controllers (for the two boards I want to use, all sata ports are connected to the built-in controllers of the CPU)

As far as I know, RAID 6 is rather I/O intensive. Say the file-system layer wants to write a 4K block. Then RAID6 requires a read of a 4K block from one disk, and one write to all other 3 disks. So in total 4 times the I/O bandwidth. Making the unrealistic assumption that I can achieve 125MB/sec through the gigabit ethernet port, it would require a 500MB/sec total throughput to 4 SATA ports. I really hope the SATA controllers are capable of this throughput.

My guess is that the computational power of the kernel's RAID6 implementation is high enough, even on this sort of Atom processor. Linux's implementation is SSE optimized and everything. I guess the same is true for FreeBSD? But I have found no results for this sort of thing. When a Linux kernel boots, or the raid6 module is loaded, it actually performs some benchmarks. Does the FreeBSD kernel do this sort of thing? It would be nice to see the results of a kernel running on such a processor. I hope, the raid6 performance is several gigabytes per sec.

Does anybody have experience with RAID6 on such a machine and could you post some benchmark results, in particular read and write performance?

 

[Ericloewe]

First of all, proper terminology is important: A group of drives with two drives' worth of parity is a RAIDZ2 vdev.

Second, Avoton Atoms are more than enough to easily saturate GbE. Do not worry about performance. In any case, the bottleneck lies mostly with the network stack, not in ZFS.

 

[skoehler]

I don't find any explanation of how RAIDZ2 works. In what sense is it different from RAID6 on 4 disks?

Do you have any numbers on the write performance of a raid6 or raidz2 on such a processor?

 

[mjws00]

Writes are NOT cpu bound. Not even a little bit. Even with a dozen ssd's you would still be bound by the drives with only a tiny bit of cpu utilization.

 

[Ericloewe]

I don't find any explanation of how RAIDZ2 works. In what sense is it different from RAID6 on 4 disks?

Do you have any numbers on the write performance of a raid6 or raidz2 on such a processor?

In principle they are the same. They do have different implementations and saying RAID6 implies a hardware RAID controller - for the sake of clarity, RAIDZ2 is the exclusive term for what ZFS does.

 

[cyberjock]

Hey,

I was going to build a RAID6 system with 4 disks based on a SuperMicro A1SRi-2750F or 2758F board. Obviously, the Atom (even though being octo-core) is not the most powerful CPU in the world. But the question is: can it handle the computations needed for reading/writing a RAID6 with 4 disks?

Absolutely. If you read up on my benchmark of the Mini with 10Gb LAN I do quite a few nasty things to my Mini to see what kind of performance you can get, and the CPU is clearly overpowered for FreeNAS unless you want to do things like run Plex for transcoding for 3+ devices while simultaneously running 10Gb on the server with gzip-9 compression and dedup. You *might* have a problem then. ;)

As far as I know, RAID 6 is rather I/O intensive. Say the file-system layer wants to write a 4K block. Then RAID6 requires a read of a 4K block from one disk, and one write to all other 3 disks. So in total 4 times the I/O bandwidth. Making the unrealistic assumption that I can achieve 125MB/sec through the gigabit ethernet port, it would require a 500MB/sec total throughput to 4 SATA ports. I really hope the SATA controllers are capable of this throughput.

No, that's not what would happen in RAIDZ2. What is likely to happen if a 4k write comes in is you'd get a 4k write to 3 disks (one with the data and two parity). Unlike RAID6, RAIDZ2 has many differences.

My guess is that the computational power of the kernel's RAID6 implementation is high enough, even on this sort of Atom processor. Linux's implementation is SSE optimized and everything. I guess the same is true for FreeBSD? But I have found no results for this sort of thing. When a Linux kernel boots, or the raid6 module is loaded, it actually performs some benchmarks. Does the FreeBSD kernel do this sort of thing? It would be nice to see the results of a kernel running on such a processor. I hope, the raid6 performance is several gigabytes per sec.

A similar rudimentary test is performed on FreeBSD. It's not overly accurate and is only meant as a roundabout benchmark. For some servers, especially big ones, there can be efficiency gains by determining exact values for your hardware and setting the sysctls/tunables as appropriate. But you will find very few people here willing to openly discuss this. It's a very detailed topic and the volunteers here are few in number so they stick to helping the masses vice the few.

Does anybody have experience with RAID6 on such a machine and could you post some benchmark results, in particular read and write performance?

Nobody here is going to have RAID6 numbers. Why? Because nobody cares about RAID6 with ZFS. Everyone uses RAIDZ2, which is not the same. If you want to read benchmarks on RAIDZ2 there's *plenty* of reviews out there via Google.