HD read/write question

[Bozo]

How does a hard drive write to a hard drive with multiple disc ( or sides)? If you have a hard drive with a single disc but using both sides: does it fill up side 1, then go to side 2; does it alternate (ie: 1-2,1-2,1-2 etc); does it split the files similar to RAID 0; or some other method. What about multiple disc?

Thanks,

Bozo :mrgrn:

 

[time]

It splits the files similar to RAID 0, i.e. after completing a track on surface 1, it switches heads to surface 2. This is called a cylinder.

 

[blakerwry]

I could also add that there's only one actuator which means all heads must be located on the same cylinder at any given time. Additionally, from what I understand, only one head may be reading or writing at a time.

 

[freeborn]

And of course it gets more complicated. On some current hard drives it's faster to stay on the same head and switch a single track then it is to switch to another head. For drives like this, some number of tracks called the segment size are written on a single head before the tracks in the same cylinders are written by another. Lets say you have 10 tracks and 4 heads with a segment size of 5. With this serpentine approach head 0 tracks 0-4 would be used then head 1 tracks 0-4 then head 2 tracks 0-4 then finally head 3 tracks 0-4 before the drive would then switch back to head 0 track 5. The sequence continues as before ending on head 3 track 9. Of course real drives have hundreds of thousands of tracks and larger segment sizes but you should get the idea.

The host never sees this though, all the host sees now is a contiguous sequence of logical blocks.


Free

 

[sechs]

freeborn -- Why would you do this?

 

[freeborn]

With todays track densities and the vagaries of head stack alignment a head switch often means a seek of around 30 tracks as well. This required seek takes a considerable amount of time. It is now much faster to stay on 1 head through multiple tracks then kick off a seek and head switch together. As long as a head switch was faster than a track to track seek, head switching was done so you read or wrote 1 track per head. Now a head switch entails a seek of around 30 tracks and so it is slower to switch heads then to seek a single track. At some point you tune the drive to determine how many tracks to stay on a head before starting a head switch / seek for optimal performance.

I saw one drive where doing the old style single track then head switch routine dropped STR by 15MB/s for each zone on a 4 head drive over the throughput provided by using the above method.

Free

 

[CityK]

On some current hard drives it's faster to stay on the same head and switch a single track then it is to switch to another head. For drives like this, some number of tracks called the segment size are written on a single head before the tracks in the same cylinders are written by another.

This sounds like it would explain what the new Samsung's are doing.

 

[sechs]

I was thinking the same thing.

I still don't understand why this would be advantageous; but I guess that's somewhere in the "vagaries."

 

[blakerwry]

On some current hard drives it's faster to stay on the same head and switch a single track then it is to switch to another head. For drives like this, some number of tracks called the segment size are written on a single head before the tracks in the same cylinders are written by another.

This sounds like it would explain what the new Samsung's are doing.

it certainly is an odd pattern.

 

[Bozo]

Thanks for the replys!

Freeborn: That doesn't make any sense at all. If you write to side one, track one and then switch to side two, track one, the heads never move. But what you are saying, on side one the heads would move into track 5 and then move the heads back to track one on side two. And that's faster??

Bozo :mrgrn:

 

[blakerwry]

if the track to track seeks are less time consuming than head switches it would be faster.

 

[freeborn]

Thanks for the replys!

Freeborn: That doesn't make any sense at all. If you write to side one, track one and then switch to side two, track one, the heads never move. But what you are saying, on side one the heads would move into track 5 and then move the heads back to track one on side two. And that's faster??

Bozo :mrgrn:

You are envisioning a perfect world. When head 0 is on track 30, head 1 could be anywhere from track 0 to track 60 in the example I was giving. The head to head offset also changes based on the temperature of the drive due to varying rates of expansion. Keep in mind that today there are around 130 thousand tracks per inch. Basically you stay on one track one head for the segment size and then kick off a seek and head switch at the same time. Once the controller is reading the destination head's servo info the seek can be adjusted to end where it is needed.

Free

 

[ddrueding]

So taking this a little deeper...

If changing tracks is faster than changing heads, why isn't it faster to run though all the tracks before changing heads?

 

[freeborn]

So taking this a little deeper...

If changing tracks is faster than changing heads, why isn't it faster to run though all the tracks before changing heads?

Outer Zones have better throughput. The best bet is to compromise, run through say 20 tracks then switch heads and seek thus staying in the better zones longer.

Free

 

[ddrueding]

Outer Zones have better throughput. The best bet is to compromise, run through say 20 tracks then switch heads and seek thus staying in the better zones longer.

Free

So it's specifically designed to have better thoughput at the beginning of the disk? Are there specific applications where this helps?

 

[time]

What Freeborn claims does make sense. If so:

A head switch is no longer the whole story; it needs to be accompanied by a small seek to re-align the head for that particular platter. So 'platter-switch' time will be a combination of head switch plus this track positioning time. If the controller already knows which direction the next track is likely to be, then movement can commence during the head switch.

The combination of 'vertical' and 'horizontal' mapping is described here.

130,000 tpi translates to a track pitch of 200 nanometers! That's smaller than a microbe and heading for virus territory. Or to compare with CPU technology, 0.2 microns. :eekers:

 

[blakerwry]

Outer Zones have better throughput. The best bet is to compromise, run through say 20 tracks then switch heads and seek thus staying in the better zones longer.

Free

So it's specifically designed to have better thoughput at the beginning of the disk? Are there specific applications where this helps?

It's just a physical limitation that the inner part of the disk will be slower than the outer.

Disk platters are shaped like records. Imagine the outter tracks of a record, they are much longer than the inner tracks. The disk spins at a constant RPM speed. This means that the outer track can be read in the same amount of time that an inner track can be read. On the outer tracks you're covering more data in the same amount of time, thus they will be faster sequentialy

Not only that, more information is held on them, so often less seeking is necessary compared to the inner tracks.

 

[ddrueding]

Blake: Yeah, I got the reson behind the outside being faster. But by doing head switches across all the heads to keep the "beginning" of the drive all in the outside is intentionally making the outside of the drive significantly faster. If you were to allow all the tracks on head 0 to be read before switching, the speed graph would have "sawtooth" marks at every head change. Weather this would contribute to a higher average transfer speed or not, I have no idea.

 

[blakerwry]

Correct. By using the fastest part of each disk at the beginning you are exacerbating the high performing beginning of the disk along with the decay at the end of the disk.

I think there have been drives like this, and I'm sure with a firmware change you could make a drive like this. A drive that performs in this manner would be great for AV use, but bad for desktop use. My reasoning is that many people only use ~20% of their disk space. Thet data is kept at the beginning, which is of course faster.

In my opinon, the end of disks were never meant to be used. Too much slowdown for seeking and for STR. But maybe that's becasue of the current design exacerbating the effects of decay.

 

[blakerwry]

just a thought: A JOBD array would perform in the same manner as what you suggested.

 

[ddrueding]

Indeed.

Don't think that I'm complaining, though...I'm only using the first 5GB of each of my raptors in RAID0

Seeks and transfer rates are pretty darn good.

 

[blakerwry]

lol. Share the wealth

 

[sechs]

So cylinders are history?

 

[freeborn]

So cylinders are history?

No, cylinders are still a valid concept to envision the physical layout of data tracks. Drives just don't write or read an entire cylinder at once now.

Free

 

[sechs]

The fact of the matter is that tracks no longer sit so neatly in alignment to think of the cylinder as geometric concept.

 

[freeborn]

The fact of the matter is that tracks no longer sit so neatly in alignment to think of the cylinder as geometric concept.

True, you would have to go back to stepper motor drives to find a drive that meets it as a geometric concept. It's a nice abstraction though that makes visualization easier.

Free

 

[CityK]

The fact of the matter is that tracks no longer sit so neatly in alignment to think of the cylinder as geometric concept.

True, you would have to go back to stepper motor drives to find a drive that meets it as a geometric concept. It's a nice abstraction though that makes visualization easier.

Uh-oh. Looks like I may have to retract my c't comment.