Teach me about dedicated Linux-based software RAID.

Order

Gawd
Joined
Dec 8, 2004
Messages
979
I've been slowly coming to the conclusion that a dedicated rack of drives connected to good HBAs and running a certain type of Linux-based RAID can vastly outperform even the most expensive hardware RAID controllers due to the fact that these drives have a whole processor to themselves for various parity/data calculations. What evidence is there to back this up? I've looked at the "TickerTAIP" design of the early 90s and it seems that if something like that was implemented today with current hardware there can be incredible performance gained.
 
The only real limitation on LSW is bus speed. I have a 3-drive raid 5 (in case you haven't heard yet ;)) using EVMS (highly recommended, btw). The processor can handle it FINE, but the fact that everything has to transit the PCI bus slows everything to a crawl.

That's sort of good news and bad news; it's slow on pci but you could get a better card with a better interface, right? Well, theoretically yes, but I haven't found one yet. The Supermicro AOC-SAT2-MV8 is an 8 port no raid sata adapter - sounds good, right? Well, linux support for the Marvell controller it uses is almost nonexistent, as far as I can tell. Your best bet may be a few sil3114 cards (impeccably supported) on seperate pci buses. I only wish SI would release an 8 port pci-X card like SM did. That would rock.

 
Sounds like a plan :D but it won't help my sata drives much, especially cause I got no pci-E...

I'd go with a card that can do some form of HW raid, just because if you then find out that LSR is too slow you can switch to hw. LSR is plenty fast in theory, but it's not a single integrated system, and the components you buy could emphasize the wrong strong points to get decent speed out of LSR. There's no guide on what parts to buy to get speed out of SW raid, which might be a good thing to come up with when I get the time and money to get a half dozen adapters and try them all. It'd make an interesting set of benchmarks, tho. SW raid would probably test the seek times of the disks and bus bandwidth the most, so it might be awfully consistent in terms of scores, but one could talk about linux support of various models, and stuff like that.

 
I'm going to do this, then.
I want to make my disk I/O system completely seperate from the compute/graphics node.
This will be a lot of fun, IMO.
 
Back
Top