MACH.2 Multi-Actuator Hard Drives | Hacker News Reader

MACH.2 Multi-Actuator Hard Drives

seagate.com

36 pointstanelpoder60 comments

Apparently both Seagate and Western Digital sell dual-actuator disks (double the IOPS and throughput up to 448 IOPS and 524 MB/s per drive)

Open articleView on HN

Apparently both Seagate and Western Digital sell dual-actuator disks (double the IOPS and throughput up to 448 IOPS and 524 MB/s per drive)

When I first heard about how mechanical hard drives work, back in the days when the highest end 300 MB (yes MB) drives ran a stack of four or six platters, and some drives still used a sort of rack and pinion positioning setup rather than a voice coil, I assumed each platter would have its own actuator and heads were only paired up top and bottom per platter.

I was surprised to find that instead, the entire rack of them moved as a single unit, with so much energy wasted overcoming 3 or 4 times the necessary inertial mass for each seek. So now, 20 something years later, well after SSDs have come close to superceding spinning rust when all tradeoffs and variables are considered together, someone has finally gotten around to putting an extra actuator in these units. Great!

While on that subject I also never understood why the general layout of components inside these drives never iterated away from the very earliest RLL / MFM winchester mechanisms, with the spinning platter assembly slightly off-center in the case, and a single head assembly off in one corner.

I assumed at some point that the armature assembly would become miniaturized just enough to allow centering the platter spindle and adding another set of heads at the opposite corner of the case, so that each head set could address opposite sides of the platter, effectively halving access times even on single platter drives.

The heads were very lightweight, and defragmentation algorithms were supposed to make sure that the heads seldom had to move far. Far more energy was wasted spinning the platters. And the benefit of the old system was that you only needed one system for ultra-precise positioning of the heads. This was probably not just simpler and more compact, but also more reliable. In a multi-actuator design, the failure of any of the independent positioning mechanisms kills the drive.

Not saying that this isn't progress, but 3.5" and 2.5" HDDs were a true marvel of engineering when they came out, and maybe we needed a couple of breakthroughs in electromechanics to get to MACH.2.

Agreed re: mechanical HDDs being engineering marvels. The positioning performance at the price/manf volumes is really incredible.

As the poster below mentions, neighboring actuator coupling/disturbance rejection is very much a performance consideration (even between neighboring HDDs) and then to package all that additional complexity into the same form factor is really something.

It'll be interesting to see if this sticks around, or the added complexity makes it short lived.

It might be difficult to control the heads if each one has other heads moving right next to it. Magnetic effects of the voice coils, mechanical vibration, maybe electrical even. It seems a much harder control problem given that the requirements are so stringent.

Rejection of neighboring HDD vibe disturbances is already an issue, so fascinating to see what their control approach is. In my limited experience, mechanical HDD control is near top of the game for electromechanical positioning.

At least they know what all the heads are doing so could design them to cooperate or model/predict the disturbances pretty well, unlike for external vibrations from neighbouring disks.

> well after SSDs have come close to superceding spinning rust when all tradeoffs and variables are considered together

SSDs are great, but I wouldn't put them that far. There are definite drawbacks to SSDs due to the nature of their design that HDDs don't have, and vice versa. Use the right tool for the job, as they say.

Personally, I have both SSDs and HDDs in my rigs because I have workloads appropriate for either of them at a given time.

I also have both, but only because I can get HDDs for $10 to $20 per TB while SSDs still cost $50 to $100 per TB.

Could you give examples of suitable workloads for HDD and SSD

Workloads that require low latency, fast random access, don't involve too much (re)writes, mobile environment, money is not an object: SSDs.

Workloads that involve lots of writes and especially rewrites, transient and ephemeral data, don't require low latency or fast random access, money is a factor: HDDs.

The key thing is that HDDs have a theoretically infinite amount of writes compared to SSDs which have a finite amount. Certain workloads are consequently going to be more suited for HDDs.

I’m not the person you answered but SSD: boot disk, cache, video files you’re working off of. HDD: data you don’t mind waiting for such as video and audio storage, backups, or just cases where you have a lot of data and the SSD isn’t yet cost effective enough

For mechanical drives, the ratio between IOPS and capacity has been getting exponentially worse over time.[1] This means that random seeks available per unit of data is getting so bad now that in cloud hosting they consider a 128 KB read to be "one" operation for cost calculations. The capacity unit for a single I/O used to be 512 bytes!

This is why I came here to make the same comment you just did.

A modern 30 TB data centre drive has about 300 IOPS, so that's just one random seek per 100 GB per second! Ouch.

I don't get why manufacturers don't make drives with actuator arms in all four corners, and heads that can move independently on both sides of every platter.

There might be some vibration and cross-talk issues, but surely they can be overcome with modern digital servo control technology.

That would allow 4x the IOPS just form having four sets of arms, and then 2x the IOPS because there's arms above and below platters, and then Nx where 'N' is the number of platters. Let's say there's 10, so that's 4x2x10 = 80x the IOPS for the same capacity!

Another thing I was always curious about is why swing-arm actuators stuck around all these years, rather than moving to a single fixed rail with a strip of individually addressable read regions, and another rail separate (and upstream) from this with write capability).

No more mechanical actuation at all, eliminating a huge amount of complicated precision machined componentry as well as the voice coil itself.

Mounting this fixed rail to multiple structural points inside the case means there would be zero possibility of a head crash.

No cross interference or mode switching between reading and writing. The entire region passing under the rail can be scanned simultaneously, or if it's not possible to manufacture such a sensor at a data density to match the drive platter then perhaps the rail could shift a millimetre or so back and forth to allow micron-fine positioning from a milimeter-course array of heads. Much less mass to move, likely simplifies the math needed to meet the data in flight, and reduces the total reciprocating moment to a single linear axis solution.

I know I can't be the only person, nor the 1st, to have considered this idea. I suspect multiple variations may be lurking in the patent portfolios of the big storage mfrs. But I would love to know why nothing resembling it has ever been tried in production over the 30 or 40 something years that hard drives have been mass commoditized.

> perhaps the rail could shift a millimetre or so back and forth to allow micron-fine positioning from a milimeter-course array of heads.

You're talking in terms of millimeters and microns. That is roughly the scale that CDs operated at.

According to https://blog.stuffedcow.net/2019/09/hard-disk-geometry-micro...

> In the newest drive I have, average track pitch is 80 nm and an average bit is 17 nm in length.

That's for hard drives providing 1TB per (double-sided) platter, less than half the density of today's most advanced hard drives.

A relative of mine designed ink-jet printer heads as wide as the paper.

Something similar could work with hard drives: an array of about 3 to 4 thin strips of silicon chips with a line of read/write heads.

They could also be staggered in tiers, one row offset from the other.

I feel like there's an alternate universe where this is exactly the design that was adopted and someone is out there suggesting our universe's setup instead, asking why not make a single read-write head that gets whipped back and forth around at mind numbing speeds seeking to different spots on the surface of the platter, and everyone's pointing out how absurdly bad and pointless an idea that would be: think about how delicate such a setup is, you want to put a flying bit of metal over a rotating metallic surface at just a microscopic distance apart - you want to rely on the Bernoulli effect to pull it into just the right proximity to the surface, and this has to work in every possible temperature, vibration, and moisture and air pressure scenario? how do you compensate for inevitable wear and tear on the mechanism in the drift and loss of precision that will induce, what about all of the fine particulates that may develop from friction over time, how do you get lubricants to last so long, all the possibilities for terminal head crashes, how slow mechanical mechanisms are versus solid state, the comparatively massive additional power draw and heat this would introduce, and so on.

Without knowing anything about the specifics of material engineering whatsoever in this domain, my assumption is that Manufacturing the read-write heads maybe one of the most expensive technical challenges and perhaps a disproportionately large factor in the cost of the entire assembly, so perhaps asking for an entire strip of these things might be like ignorantly asking, if the light-refracting feature of a diamond is the most salient feature of a diamond ring, why not make a rings with diamonds all the way around instead of just one.

You're technically correct but I think rather missing the point. I got the scale wrong (guessed at the tolerances involved rather than looking them up), but obviously resolving a mechanical position to meet the actual tolerances is already a solved problem by current voice coil systems. A simplified version of the same mechanism could be applied to the line-of-heads solution I'm suggesting, considering that in both cases a comparable reciprocal mass is being moved back and forth, but in the scenario I'm proposing it would only need to be moved a millimeter (or two, or one, or a micron - all are far smaller of a movement than the comparatively huge arc swept by current head designs)

Whatever subdivisions within that bounding range are necessary to achieve equivalent (or superior) accuracy to flying head designs has already been proven to work in spite of the current designs' much more complicated and delicate arrangement.

Another way to look at is that the ratio between capacity and physical IOPS constraints have gotten exponentially better over time. It's just that you can't treat all that capacity as hot capacity anymore.

I recall a whitepaper on HP high-end storage arrays from over 10 years ago, their arrays already read and cached at least 256kB of data, even if you wanted only a 4kB block.

It's this just a response to the narrowing of the use cases for mechanical drives, so they specialize at long term, infrequent access, large storage?

What I do not understand yet since the floppy drive days is this:

If HD controllers are blinding ASIC-fast, why can’t we have three 120-degree actuator arms, or 4 arms, 5?. Or even a radial axis of 120 piston-like actuator arms full of magnetic read heads.

We have obviously nailed the head-of-line queueing, the quality-of-service queue, the anti-quantum tracking guidepost, the ability for N-head on same track, the constant density per orbital track, the buffeting airflow magnetic head reader (via semi-vacuum enclosure), the thermal drift/expansion, and the anti-shock advantage of its radial piston arm’s magnetic read head not “buying the farm” over sweep arm.

It isn’t about the sweeping arc of the actuator arm anymore. You can now have piston-like (albiet slower) actuator arms but the sheer number of radial axis piston-actuator should make for faster read retrieval and maybe write as well … than SSD?

Heck, we make printer heads with interval thinner than its paper, 1,024 Cochlear implant point probes over millimeter span, surely a multiple non-moving radial axis of flying scanning multi-million read heads can be floating across the entire platter? (A fixed platter silicon chip of read heads floating on top of a spinning magnetic platter?)

Of course, this is just me, the armchair quarterback of floppy drives and thinker of cross-domain technology. Please be kind.

I realise the 'MACH' name here doesn't seem to be directly related to mechanical speed, but back in the 90s I remember seeing a quote in a computer magazine about hard drive head seeking and reading, with something along the lines of it being equivalent to: "flying in a fighter jet 20 feet above the ground at Mach 2 and counting the blades of grass as you go by"

I wonder how accurate that is as a comparison (both now and 30 years ago): I'm guessing there was a fair bit of exaggeration going on, but maybe not.

Well, I'm stuck in an urgent care waiting room, so let's give this a shot.

Mach 2 at sea level: 1,522 mph, or 2,232 feet per second.

1 square foot of lawn has around 3,000 blades of grass.

2,232 square feet * 3,000 = 6,696,000 bps (blades per second)

Hard drive transfer rate around 30 years ago: 80,000,000 bps (bits per second)

So honestly, closer than I expected! Hack the width of the grass strip a bit and you'll get a perfect match.

Is it wrong to say that this is mostly just two drives stacked on top of each other, inside the same enclosure? If so I'm not quite seeing the benefit over RAID 0.

I was expecting multi-actuator to mean two sets of heads sharing the same platters. Now THAT would be a fun engineering challenge.

Sort of.

As I understand it: By itself, it's theoretically about the same as two (non-RAID) hard drives. They show up as separate LUNs [or similar]. Like two standalone drives, these can be used independently or made into RAID 0 or lots of other things.

But compared to using two drives, these have half as many spindles, half as many spindle motors, and half as many spindle bearings.

That's not a small reduction in complexity, and reductions in complexity tend to improve MTBF.

Furthermore, they can theoretically use about half of the physical volume. And half the external connections and cabling.

And (maybe, some day) this can all combine to reduce the overall cost of adding a headstack (or a thousand of them) in some applications where that is a useful thing.

Huh. I assumed that it would have two columns of heads on opposing corners. Twice the throughput, half the seek time.

Instead it still only has one column, but it’s divided in vertically to give more flexibility in positioning. I understand how that could help with random access iops, but how does it double the max bandwidth as they claim?

It’s still the same number of heads isn’t? When doing sequential reads isn’t that the limiting factor?

I guess I've always assumed that HDs wrote with just a single head: moving to a specific cylinder, using a specific head, and then writing the appropriate sector (CHS). But, I guess it would theoretically no different to write the stripe across all the heads in that cylinder, though for efficiency maybe the sector size would have to expand by NR_HEADS times?

I guess this goes back to various diagrams I've seen over the years.

Anyone able to enlighten us?

Hard drive tracks are really narrow. When you have one head aligned with the track on one platter, the rest of the heads on other platters will only be close to the corresponding tracks on those platters, not fully aligned.

Now that hard drives are using multi-stage actuators (arms with elbows and wrists), it's theoretically possible to align multiple heads to allow writes to be striped across multiple cylinders. But I suspect those advances have instead been used to enable even tighter track spacing (and overlapping, for SMR drives) because SSDs have forced hard drive manufacturers to prioritize density over performance.

Are tracks physical things in modern disks?

I’ve long assumed we got to the point where the surface of the drive was uniform recording material, “tracks” were just the parallel bands the disk made by recording data. There’s difference in the platter material between two tracks, just the gap the heads leave to avoid interference. Conceptually like audio cassette tape: it’s all magnetic, but the audio is stored in separate tracks.

Is that wrong? Is there some physical coating or gap in the recording media between tracks?

If that’s not wrong, wouldn’t always reading/writing the same track across multiple heads force them to stay in alignment?

I guess I always assumed all the heads were independent data streams, they were just forced to move physically in parallel.

You’re right that reading or writing the data in parallel across multiple platters might make more sense since the heads are always locked together anyway.

I have no idea how it actually works. That had never occurred to me.

In a traditional drive, only one head is active at a time. In these drives only two heads are active. I don't think multiple heads on the same arm can be used at the same time due to micro-misalignment of the tracks.

Well that would explain it. I’ve been trying to Google it and it and everything I’ve found matches what you’re saying. Thank you.

You’d have to make sure that your data is always on different sets of platters to take advantage of the double throughput though. You could get 2x throughput if it was all on platter 1.

With the misalignment issue you mention exist if we were using multiple heads? If things were ALWAYS done in parallel across the heads didn’t wouldn’t the tracks always have to have the same micro alignment as well? Wouldn’t the problem only exist if a track on platter A and a track on platter B were written separately?

Tracks are now so narrow that thermal expansion might throw the alignment out of whack.

IIRC, hard drives have had the problem of thermal expansion vs track density for decades now.

I helped build a small recording studio around the turn of the century, and at that time "AV rated" drives were a thing that demanded a premium price.

These differed from other drives in that they were capable of continuous use, whereas other drives might briefly go out to lunch periodically to re-do their thermal calibration.

This was considered important by some, since it was critical that the process not "miss" any data due to the hard drive being late to the party during any part of a recording or mixdown session. These were once new potential problems in the recording space (previously, we used tape -- a linear medium for a linear event).

We don't talk about tcal these days and terms like "AV ready" have dropped from the data storage lexicon.

I can only assume that it ceased to be a practical problem somehow. Maybe tcal happens fast enough now that any writes can be caught in cache, or maybe tcal happens invisibly as a continuous process, or maybe data rates for spinning rust have improved enough that we're no longer so close to the edge of hardware performance that it ever matters, and/or maybe operating systems and recording software have improved enough that it's just not an issue worth discussion.

(These days, cutting a music track of ridiculous complexity on a singular MacBook is a complete snoozefest for the hardware. But it hasn't always been this way...)

I assumed that too. Probably cheaper and more space efficient to have just one actuator position in the HDD enclosure though and they split it vertically.

They keep saying it doubles the iops, but is that really accurate? From the diagram it seems like half of the platters have their own actuator. Surely unless the data is physically distributed across the platters in a really funky way you'd only see a doubling of iops when accessing two different pieces of data that are each found on platters addressed by their own actuators.

I was thinking that the controller probably takes care of interleaving data across platters, so at least small enough random I/Os would be able to use only one actuator.

I did some more reading and the PDF [1] says it shows up as two separate LUNs behind a single SAS port, but sharing the drive cache.

But what I initially didn't catch (and another comment here pointed out), is how come they claim to do over 500 MB/s of sequential scanning I/O. Small random IOPS I get, but for large sequential I/Os you still have the same amount of platters, with the same bit-density, rotating at the same speed... Or is even the sequential I/O throughput "bottleneck" mostly dominated by the tiny seeks + calibration to the next adjacent track and not the actual magnetic reading part?

I now really want to buy one for (some?) reason. Could get an used version from amazon for $150 but I'll try to hold back. It's also interesting that there are no new, unused versions available there (discontinued?)

[1] https://www.seagate.com/files/www-content/solutions/mach-2-m...

Transfer rates can be doubled by spreading (something like) RAID 0 across those two LUNs, if doubling transfer rate and treating it all as one logical device is the goal.

The drive doesn't know how to do this by itself.

(That doesn't mean that the marketing wank is a lie: It's still a singular drive with a singular SKU. The marketing wank may easily be considered to be incomplete, however.)

These drives have been available for a fair time now, though they're still fairly new. They were never intended for plebs like us to be able to buy from places like Amazon, probably because the support costs would be through the roof for such an unusual product with such niche use-cases.

They probably stripe the data across the banks of platters just like multiple drives in raid 0, so when reading or writing a sufficiently large file, both banks are always operating at full speed.

But if the number of platters, rotation speed and bit-density is exactly the same as in regular drives, the number of “bits” moving past the read/write heads per rotation would still stay the same? Would be interesting to know how much time typically is spent on actually reading the bits vs the tiny seeks & track calibration when reading, say, 10 GB sequentially.

Edit: other comments below explain the missing part, in a regular drive, apparently just one head can read/write (to one platter) at a time due to calibration reasons, even if your disk physically has 8 platters. So having two actuators allows two platters out of 8 to do IO concurrently.

I remember (many) years ago seeing an IBM mainframe storage drum.

It had a few platters, each with 4 heads every 90 degrees, and each heads was the width of the platter. The fixed heads had one read/write head over each track.

sort of brutally effective.

EDIT: found it. IBM 2305 fixed head storage module:

https://ed-thelen.org/comp-hist/IBM-ProdAnn/2305.pdf

https://bitsavers.org/pdf/ibm/2835/GA26-1589-5_2835_2305_Ref...

poking around looking for this, it seems that lots of early IBM disk drives had multiple heads.

They would have some conventional heads, and a few fixed heads over tracks with frequently accessed data.

I could have sworn that back in the '90s I heard about some enterprise drives that had multiple sets of actuators to give multiple heads per platter. However, in searching around I'm not finding any references to any. Maybe I was just daydreaming about "that would be cool", but in reality it just made more sense to throw more drives at the problem and get more parallelism through additional drives. Maybe the platters weren't the bulk of the cost, perhaps even the actuators and head assemblies were the bulk of the cost, making it totally not make sense (IOW, the drive wouldn't be marginally more expensive, it would be twice or more as expensive).

Looks like WD has multi-actuator drives as well: https://www.techtarget.com/searchstorage/news/365530452/West...

> multiple heads per platter

Conner Peripherals "Chinook": https://en.wikipedia.org/wiki/File:Conner_Peripherals_"Chino...

https://en.wikipedia.org/wiki/Conner_Peripherals

It was apparently less of an innovation and more of a desperate measure to offset numerous other technological shortcomings.

That sounds vaguely familar, thanks for that pointer.

I built my most-recent ZFS pool on 2x18 drives, mirroring pairs "crossed" across drives. The SAS models are easier to work with because they present as two LUNs. Wendell from Level1Techs wrote some scripts (easy to find on the L1T forum) to make it easier to split up the SATA models into two partitions that can then be used as mostly-independent "logical units."

In general I've been happy with it but buying the drives new probably wasn't worth the cost. ServerPartDeals has refurbished SATA drives for $210 and that's an attractive price. It's great for smaller deployments where you might have a limited number of 3.5" drive slots available but need to hit certain sequential throughout and IOPS numbers.

Of course used enterprise SSDs are probably a better choice if you can find a good deal and buy extra redundancy.

That's great, now add controllers so that you can read/write all of the heads on an actuator in parallel... Won't help the IOPS, but it would shoot the throughput up by an order of magnitude in a multi platter drive.

This was alluded to in one of the Sun performance "bibles" in the section on drive IO ... the book spoke of some "non commodity disks" that had multiple disk arms that moved independently on different platters.

I think it was "Sun Performance and Tuning: Sparc & Solaris" (the Porsche book) but it might have been "Configuration and Capacity Planning for Solaris Servers."

Cool tech. The drive manufacturers are essentially marching to the drum detailed in this paper from 2016: https://research.google/pubs/disks-for-data-centers/. Multi actuators are part of it but also higher density and better tail latency.

So it’s just one stacked on top of the other. Presumably, then, the competitor product is two hard drives. Is this /much/ cheaper? Presumably better for rack space. Power consumption? Does it even matter that it uses one fewer connector?

It’s going to be a hell of a time telling the file system to stripe across drives for redundancy but not /these two/.

I imagine if you stripe across multiple drives that are all actually smaller striped arrays internally but each emulating single drives with firmware, you still get multiplicative speed increases, e.g. a raid 0 with two of these is 4x as fast, and still just raid 0 to the raid controller. I’m not really sure how that is actually better than twice as many regular drives either.

Why would you assume the disk is presenting itself as two drives rather than one and handling the logic internally?

I stand corrected. The FAQ mentions appearing as two LUN's but only in the SAS version.

https://www.seagate.com/content/dam/seagate/migrated-assets/...

Are these the ones that appear as two volumes? These have double the risk of failure, so you really have to run them in RAID1, halving the IOPS and storage. You get no benefit aside from lower power consumption and denser RAID boxes, but have to be careful because RAID on two volumes in the same shell is risky so you should RAID across shells.

Do they really have double the risk of failure as a combined unit? I'm not sure I see the actuators fail very often in HDDs.

This only increases speed if the data you want is on the opposite side of the drive.

That's a good point, as if you don't seek, the total number of disk platters, rotation speed and bit-density is still the same regardless of two actuators. (Or do they have more platters?) Makes me wonder how come they advertise 500+ MB/s sequential IO rate then.

They literally and physically doubling down on spin disks when solid state is a about to eliminate them

"Fuck Everything, We're Doing Five Heads"

"Seventy-six cutting edges … wrap around an entire face and shave it close and clean in two or three effortless moves."

https://www.wired.com/2004/03/stark-shaving-mad-dept/

See also the Seagate Elite 47 from 1997? 5.25" full height, 14 platters, 28 heads, wide scsi-2?, 47GB [1]

Of course, 14 platters doesn't sound like that many anymore. High capacity 3.5" drives now use nine platters [2], although it's harder to find the specs.

[1] specs burried in the user manual https://www.seagate.com/staticfiles/support/disc/iguides/scs...

[2] https://www.tomshardware.com/reviews/seagate-ironwolf-pro-20...

Reply on news.ycombinator.com