[Arm-netbook] EOMA server standard

[Gordan Bobic]

On 10/26/2012 05:11 PM, Mehmet Mersin wrote:
> On Sat, Oct 27, 2012 at 12:45 AM, Gordan Bobic<gordan at bobich.net>  wrote:
>> On 10/26/2012 04:29 PM, Mehmet Mersin wrote:
>>
>>> My reasoning is even for a home server, if it's a file server I want to have
>>> at least 2 SATA ports for RAID-1. And for a router I want to have at least
>>> 2 native (not USB) ethernet ports. But these also mean EOMA-68 is not for me :)
>>
>> Why not instead simply have a SATA port expander on the chassis side and
>> run it off the single on-board SATA port? There's nothing to stop you
>> from plugging a 5:1 SiliconImage SATA expander into the card's SATA
>> port, is there?
>
> I heard (but haven't researched much or tried)

I did. And it's not that big a deal. You have far bigger problems than 
SATA port multipler bandwidth:
1) CPU's checksumming performance (does it have a built in xor engine 
like the Marvell Kirkwood?)
2) Network bandwidth (how much do you think you'll plausibly get out of 
your gigabit port?)
3) How much CPU do you think you'll need to run Samba/NFSD/iSCSId in 
addition to RAID checksumming? Will this run out before your disk bandwidth?
4) Is your load linear or random? If you are using a 5-port SATA 
expander like the Silicon Image SIL3826, and have 5 disks plugged into 
it churning out a typical 120 IOPS (assuming 7200rpm), that's 600 IOPS 
in total, peak. Assuming these are 4KB IOPS, you are looking at 2.5MB/s. 
You aren't going to bottlenecked on your SATA port if you are IOPS limited.

> that raid performance
> of port expanders, especially cheap ones are not very good.

Depends on the host controller. If it supports FIS switching, you'll get 
more bandwidth right up to the point where the controller runs out of 
bandwidth. Without FIS switching support, the performance will tend to 
drop off slowly as you add more disks, but the chances are that it'll 
still easily saturate what you'll get from your gigabit network port.

> There's no bandwidth problem
> with single SATA-3 port
> having 600MB/s theoretical limit, but I don't know any decently priced
> expander chip
> that can take advantage of this bandwidth. Also not all SATA
> controllers in SoC's
> support expanders.

How much of this is due to driver/firmware bugs and bad design causing 
race conditions? In theory, all SATA controllers should support port 
multipliers. The good ones will also support FIS switching and thus have 
better performance.

> And when you use RAID mode of these chips and show
> your RAID setup
> as a single drive, performance is worse than soft RAID.

That's debatable because you only have an ARM SoC to be doing your RAID 
checksumming, not a big fat 64-but x86 CPU.

Gordan