Why are multi-platter drives so slow?

Tomsk

Tomsk

Tomsk

Soldato
Joined
18 Oct 2002
Posts
5,236
Location
Overground, underground..
One for the HD boffins.

If you look at the benchmarks in this post - http://forums.overclockers.co.uk/showpost.php?p=11522718&postcount=10 - you can see that the multi-platter drives are of similar speed to the single platter drive.
From these benchmarks one can conclude that the drives are reading off only one platter at a time.

What prevents a drive reading from more than one platter simultaneously?

It seems to me that a 2-platter drives could be twice as fast as a single platter and a 3-platter drive could be three times the speed of a single platter.
 
something ive often wondered. a sort of internal raid set up. i asked this question one and i was only told 'they already work like that'. i cant say im convinced. if a single platter will achieve 80mb/sec average, say, the 3 of them in one drive should achieve close to 220-230mb/sec. but they dont......
 
I've wondered this for years and never come across any techinical explaination of why they don't do this, despite a fair few people labeling my suggestion as impossible/crazy. It should also scale perfectly, unlike RAID-0 across multiple devices - as they have to deal with slight rotational latency differences - which could be helped by having HD's continuously read the track under the head into a circular buffer - don't know why they don't do that either :p
 
Last edited:
I'm trying to think of what would need to change to make this happen. So far I've come to the conclusion that you'd need to completely re-jig how disks are accessed, most likely through some kind of request remapping on the drive itself.

At the moment data is accessed by block (although sector is a more common, if confusing name) address which is a unique combination of the track (radial position), sector (linear position within the track) and head (top or bottom of which platter).

To get data off the platters simultaneously the block needs to spread across all platters in a column so that it is addressed simply by track and sector. The disk then just pretends to be a single headed affair and deals with the block internally. Reading and writing should be reasonably straightforward although there would need to be some buffering involved to convert the serial data stream to and from the PC into a 2, 4,6,8 or 10 bit parallel stream to and from the drive heads.

There are implications to be considered though:

  • How do you deal with a bad block? You'd need to remap the entire column rather than just a block on a single side.
  • Assuming you vary the block size as you add more platters what happens to files smaller than the block size? How do you use the rest of the block?
  • Is it possible to do the serial to parallel conversions quickly enough to cope with the data rate coming off the drive?
 
The drive would just take the LBA address and take the modulo vs the number of heads, ie for 3 platters: LBA sector 1000 mod 3 = head 1, 1001 mod 3 = head 2, 1002 mod 3 = head 0. each head can read every sector under it as the disk rotates into a buffer and thus gets 3 sectors on every sector that is rotated.
Eg, if the host asks for sector 999 (head 0), then sectors 1000 and 1001 will also be in the exact same place at the same time and can be buffered, prepared for if the host asks for those (likely, in reading a contiguous file).

I don't see why an entire cylinder would have to be relocated on a bad sector, just relocate normally, the other sectors could stay at the same place. Of course that means that the entire actuator arm has to move all heads at the same time to read the relocated sector and abort buffering the request it was performing from the other heads.

I think the processing requirements are well within the realm of an off-the-shelf Xscale CPU or ASIC.

Perhaps the real problem is patent issues :

http://www.patentstorm.us/patents/7324301.html
 
Last edited:
matja said:
Perhaps the real problem is patent issues :

http://www.patentstorm.us/patents/7324301.html
Click to expand...

I never envisgaed the actual arrangement of data on the platters to be a problem.

From the patent application it seems the problem is the mechanical alignment of the tracks. The patent gets around this by having each head controlled by a seperate positioner.

That patent was only granted in January this year, in the time it takes for any manufacturer to develop a viable commercial product, SSDs will have probably gained a significant proportion of HD sales to make the technology non-profitable.
 
The read head is fixed between all drives so it can't move to two places at the same time, for that to work you need a read arm per platter and it might not be possible to fit that in a 3.5" design :confused:
 
Dutch Guy said:
The read head is fixed between all drives so it can't move to two places at the same time, for that to work you need a read arm per platter and it might not be possible to fit that in a 3.5" design :confused:
Click to expand...

ah bit what if it read from all platters simultaneously, arranging each next logical sector on the next platter?
 
You must log in or register to reply here.
Back
Top Bottom