TR Forums

we all know hard drives have RAID configurations....

what if you could do the same for memory? have two 512 sticks of PC3200 act as one 512 6400, or DDR 800.... P4c....

good idea? bad idea? too hard to get to work?

Well if you had two 512MB sticks of PC3200 running in "RAID", you would have 1GB of PC6400, not 512MB. So what you're talking about is pretty much dual channel, since you would need two memory controllers anyways.

crazybus wrote:

Well if you had two 512MB sticks of PC3200 running in "RAID", you would have 1GB of PC6400, not 512MB. So what you're talking about is pretty much dual channel, since you would need two memory controllers anyways.

no, I mean put the same data on each of them, and be able to take it twice as fast.

even with dual channel, unless you put the same data on each of them, or find a way to perfectly spread the data on the two chips, you're not going to see anywhere near double performance

If you put the same data on both of them it would be running like a RAID 1 setup and give you no performance increase.

crazybus wrote:

If you put the same data on both of them it would be running like a RAID 1 setup and give you no performance increase.

you could take data off of them faster, because you would double your bandwidth to ANY piece of information

crazybus wrote:

If you put the same data on both of them it would be running like a RAID 1 setup and give you no performance increase.

RAID 1 does increase read speed. It's write speeds that are lower (same data has to get written in 2 devices, which will give you the speed of the device at best, with a bit of overhead for synchronization).

RAID 0-like dual channel does much more sense than that. ECC would work better if you look for safe mem.

I don't see why it would be any different than a RAID 1 setup. The information still has to be read from each stick. Think about it this way:

You have 1MB of data that is saved in memory. To double bandwith half of the file would have to be read from one stick while the other half is read from the other stick. In that scenario there would be no use in have more than one half of the data on each stick, since it would not be used anyways. If all the data was read from each stick it wouldnt be any different than just have one stick, unless perhaps there would be latency benefits like you see in a RAID 1 array. Ok that was kind of confusing and I probably got it all wrong.

Crazybus has it correct.

Dual channel = RAID0.

Ram access the way you're thinking would not be any faster, it'd be very similar to RAID1.

Two 40GB hard disks in RAID0 = 80GB. Two 40Gs in RAID1 = 40GB.

Two 256 3200 sticks in dual channel = 512 6400, two in this mirrored RAM mode would be 256 PC3200.


And dual channel can get you *very* close to doubled bandwidth, I hold my i875P as proof of that.

Forge wrote:

Crazybus has it correct.

Dual channel = RAID0.

Ram access the way you're thinking would not be any faster, it'd be very similar to RAID1.

Two 40GB hard disks in RAID0 = 80GB. Two 40Gs in RAID1 = 40GB.

Two 256 3200 sticks in dual channel = 512 6400, two in this mirrored RAM mode would be 256 PC3200.


And dual channel can get you *very* close to doubled bandwidth, I hold my i875P as proof of that.

how good does that dual memory channel work? does it split the data across both sticks?

I think it is called interleaving

crazybus wrote:

I don't see why it would be any different than a RAID 1 setup. The information still has to be read from each stick. Think about it this way:

You have 1MB of data that is saved in memory. To double bandwith half of the file would have to be read from one stick while the other half is read from the other stick. In that scenario there would be no use in have more than one half of the data on each stick, since it would not be used anyways. If all the data was read from each stick it wouldnt be any different than just have one stick, unless perhaps there would be latency benefits like you see in a RAID 1 array. Ok that was kind of confusing and I probably got it all wrong.

You don't have to read the same thing twice in RAID 1. You have to read half from a source and half from the other source (assuming a typical 2-drives configuration). Good RAID-1 setups read faster, but write slower (synchro overhead) than single drives. They do read different things from both drives simultaneously.

What's the use for having the data repeated? in HDs it's redundancy. If a drive goes South, you can replace even without turning off your computer (hot-swap). I don't know of any RAM that allows hot-swapping, so RAID-1ish configurations don't make any sense. Dual channel does.

IMO, RAID-5 and RAID-6 are better solutions than RAID-1 when both performance and redundancy are critical. But those need more HDs and are much more expensive.

You shouldnt even try to apply raid to RAM, unless you use a RAIDed ram drive, and that will suck memory.

Interleaving has to do with how many banks that can be accessed in turn without refreshing(which takes time) a module, normally you have a request, then a refresh, interleaving takes several requests to different banks and interleave then, getting rid of a lot of unnecessary refreshes in between. And as each module usually have two banks or more, its a way to speed memory access up a bit. Still bound by the single channgel bandwidth restraints though. Dual channel works in a different manner where you have dual physical paths to the northbridge, being able to transfer more data at any one time. This can probably be combined with interleaving having more banks open at the same time.

how good does that dual memory channel work? does it split the data across both sticks?

Quite well. I get ~5.5GB/s in Sandra's mem bench, and it shows in many real world tasks (DivX encoding) as well. I got about 2X more performance in DivX encoding from my 2.4C+dual DDR400 than I got from either my 2.4B or my 2400+. It was scary fast.

Even memtest86 is faster. I can finish a full pass in ~10 minutes, even though it doesn't ID the chipset correctly (correct chipset support roughly doubles memtest's speed). The same pass, run on an unidentified nForce2, takes almost half an hour to complete, and that's with 1/2 the amount of ram.

If your app is memory bandwidth sensitive, i875P and a C P4 is impossible to beat. It even spanks PC1066/PC4200 RDRAM, in bandwidth and latency both.

This is sort of related, but not entirely.

Why not create RAIP? Ok, the acronym is horrible, but why not create a drive with two independent head actuators that read/write to two different platters with striped data? Since the problem with drives is not the interface speed, but the speed of actually getting the data off the platter, wouldn't this be a viable solution?

I know that there would be some interesting design issues to come over, namely making a drive with two independent head actuators and the electronics to control them, but such drives have been created in the past. Connor Periphals(which no longer exists since Seagate bought them years ago), created a drive which had two actuators.

I guess the real reason this has never come about is simply because it's just cheaper to buy two drives and do RAID, plus you have more redundancy that way.

I think the idea is still interesting though, even if not really practical because of economic reasons.

Answered your own question.

Cheaper and more redundant, to do RAID.

i don't think setting up your RAM in a RAID 1 fashion would do anything. RAID 1 reads faster and writes slower. it reads faster because it uses whichever drive finds your data first. so you minimize your seek time. it writes slower because you have to write to both. so the amount of time is the time of the drive that takes the longest. AFAIK RAM can access any portion of its memory in the same amount of time. so there is no variable seek time and therefore no performance increase or loss except for system overhead.

i'll agree with muyuubyou. using the extra drive for parity bits is the way to go. almost as fast as RAID 0 and almost as safe as RAID 1. and the more drives you have the lower the proportional storage loss. because you only ever need the one drive for parity.