Suggestion: Disks Option, add Spin-Up delay to reduce peek current

[troun]

In disk option, where smart options, APM and AAM can be managed, it will be a nice feature to add the possibility of a Spin-Up tempo/delay.
Hard drive when they start/spin-up can consume very high current for a short time (2, 3 or even 4 times the normal duty/working current), and when you multiply by the number of disk in the array, you quickly arrive to significant peek power that can bring instability and push to build over-dimensioned power supply.
Adding a delay may allow to spread this peek current over a few second, with a fairly acceptable trade-off.


Just for example, a few 3,5" recent hard drives but also valid for 2,5"

 

[Ericloewe]

This is not something the OS can do much about. Disks spin up way before the OS is loaded.

 

[troun]

This is not something the OS can do much about. Disks spin up way before the OS is loaded.

Sorry if I was unclear, but I mean after the hard drives have been in standby. When they are "waked up".
I am not talking of startup.

 

[Bidule0hm]

Usually you don't want to spin down the drives on a NAS so this feature would be barely used.

 

[troun]

Usually you don't want to spin down the drives on a NAS so this feature would be barely used.

Why Usually? From which statistic comes the "barely used"?

I personally think the contrary, power consumption optimisation on NAS, imho, makes sense.

 

[zambanini]

the power spike do not mean it will use less power overall if the disks spin up delayed.

 

[Bidule0hm]

Why Usually? From which statistic comes the "barely used"?

I personally think the contrary, power consumption optimisation on NAS, imho, makes sense.

When you spin up/down the drives you add wear to them, so you'll need to change the drives more often (and that cost money), so actually you've achieved nothing but replacing the drives more often.

 

[troun]

the power spike do not mean it will use less power overall if the disks spin up delayed.

For sure, the total amount of energy used (power over time) will be the same, just it may allow to spread the peak over time.

Let's take an example (the M9T I gave in first post, assuming 1.5 sec spin-up) to make it you easy:

You are at work, your NAS (with 8x M9T RAIDZ2) is idle for 8 hours, and you want to watch a movie, so you wake it up through your Raspberry Pi XBMC.

_Current case; all hardrives wake up together, so you get a power peak of = 8 * (4.3W-0.87W) = 8 * 3.43 = 27.44W in 1.5 second

_ With delay case: You wake up your hardrives 2 by 2 = 2 * (4.3W-0.87W) = 2 * 3.43 = 6.86W in 1.5 second, then again 6.86W in 1.5 second, then again 6.86W in 1.5 second, and finally then again 6.86W in 1.5 second.

So energy is the same 27.44W over 1.5 second and/or 6.86W over 6 seconds. You just trade off time for lower Power peak.

 

[troun]

When you spin up/down the drives you add wear to them, so you'll need to change the drives more often (and that cost money), so actually you've achieved nothing but replacing the drives more often.

it depends on every ones use, with properly configured spin-down (not 20 seconds like on some Western Digital Green, but 20 min), you achieve 1 to 2 spin-down per days, (but let assume 10 load/unload cycles), with 200/300k specified harddrives, it gives you 55 years before reaching it (worst case).


Edit: on same basis, with just ~1W saved per drive, let's say just for 8 hours (more likely to be 20 hours per days with an average of 4 hours of NAS domestics use), over 1 year, it makes almost 3kWh saved.
Multiply by how many hard drives you have (your case is 8?) over the 55 years of LCC lifespan...

 

[Bidule0hm]

You've mixed up two very different things:

• Head parking (and the associated LCC value)
• Drive spin down

The first is not harmful (unless the timer is really low like the 5 sec on greens) but the second is.

 

[troun]

You've mixed up two very different things:

• Head parking (and the associated LCC value)
• Drive spin down

The first is not harmful (unless the timer is really low like the 5 sec on greens) but the second is.

I don't mix anything, spin down and head parking are getting always together.

_ if you spin down without head parking, heads may touch platters surface and damage it. The turning is absolutely necessary to create a layer of gaz (air or recently helium) to prevent contact between head and platters.
_ Parking heads without spin-down doesn't make absolutely any sense! It doesn't save energy, it just doesn't bring anything (except increasing LCC as you mention)

 

[zambanini]

troun I do not see it as a real world problem. you will need to supply the power an a boot process anyway. so even if you have a limited power source like a batterie for example... it will not help. if power is really a problem then you would stop using ecc ram. and then FN is the wrong appliance anyway. the start/stop cycles should not grow unnecessary.

 

[Ericloewe]

_ Parking heads without spin-down doesn't make absolutely any sense! It doesn't save energy, it just doesn't bring anything (except increasing LCC as you mention)

Yes, it does. By removing an obstacle (the heads), the air moves more freely, exerting a smaller shear stress on the platters, which finally leads to reduced torque requirements to maintain angular velocity.

No drive I've seen takes it upon itself to spin down without an explicit command from the host.

 

[troun]

troun I do not see it as a real world problem. you will need to supply the power an a boot process anyway. so even if you have a limited power source like a batterie for example... it will not help. if power is really a problem then you would stop using ecc ram. and then FN is the wrong appliance anyway. the start/stop cycles should not grow unnecessary.

I dont understand what you mean with ECC ram and what is FN.
For the boot up, having a voltage drop is not an issue, while on duty it is most likely to be.
It is a real world problem, believe me :)

 

[Bidule0hm]

I don't mix anything, spin down and head parking are getting always together.

_ if you spin down without head parking, heads may touch platters surface and damage it. The turning is absolutely necessary to create a layer of gaz (air or recently helium) to prevent contact between head and platters.
_ Parking heads without spin-down doesn't make absolutely any sense! It doesn't save energy, it just doesn't bring anything (except increasing LCC as you mention)

No, you can park heads without spinning the drive down.

- Yes.
- Wrong, it saves power (less drag on the platters + heads actuator and his control logic shut off) and it can avoid head crashes.

 

[troun]

Yes, it does. By removing an obstacle (the heads), the air moves more freely, exerting a smaller shear stress on the platters, which finally leads to reduced torque requirements to maintain angular velocity.

No drive I've seen takes it upon itself to spin down without an explicit command from the host.

seriously dude? Most of energy is taken by the motor with the rotation. But ok if you want.

No drive I have taken was parking heads without spinning down... Check thread with guys that disable freenas idle, but still had huge LCC on WD hard drives if you don't believe me.

 

[troun]

No, you can park heads without spinning the drive down.

- Yes.
- Wrong, it saves power (less drag on the platters + heads actuator and his control logic shut off) and it can avoid head crashes.

Yes... I got it, huge saving over power drawn by a motor spinning 5 to 10k... Whatever I may also speak of influence of cosmic rays...
My bad.

 

[jgreco]

There's a misconception here, and it's pretty fundamental. Properly sizing your power supply IS NOT "OVERDIMENSIONING". It is properly sizing. It's like any other engineering discipline, you do not design for the average use, you design for the worst case.

The real problem here is that delayed start is a very poor option. First off, it has to be configured correctly to not start the spindles when the system powers on; this is complicated. Next, the fundamental problem is that there's no instrumentation inside a UNIX OS to create dependencies between disk devices; you'd have to have some understanding of what drives were on which power supplies, and then suppress all I/O to spun-down drives until some sort of delay parameter had allowed the previous disk to spin up. On a large 11-disk RAIDZ3, with five seconds allowed for a disk to spin and stabilize, that means I/O could cease entirely for almost as long as a minute, which is unacceptable behaviour in a filer.

 

[Ericloewe]

seriously dude? Most of energy is taken by the motor with the rotation. But ok if you want.

In steady-state, the motor only has to overcome mechanical drag. This comes from the bearings keeping everything in place and from the air, which has a certain viscosity.

Don't dismiss the latter outright - it's significant enough that it's now the limiting factor for how many platters you can cram into a drive, hence why manufacturers now use sealed drives filled with helium for the ultra-high density storage market.

 

[troun]

There's a misconception here, and it's pretty fundamental. Properly sizing your power supply IS NOT "OVERDIMENSIONING". It is properly sizing. It's like any other engineering discipline, you do not design for the average use, you design for the worst case.

The real problem here is that delayed start is a very poor option. First off, it has to be configured correctly to not start the spindles when the system powers on; this is complicated. Next, the fundamental problem is that there's no instrumentation inside a UNIX OS to create dependencies between disk devices; you'd have to have some understanding of what drives were on which power supplies, and then suppress all I/O to spun-down drives until some sort of delay parameter had allowed the previous disk to spin up. On a large 11-disk RAIDZ3, with five seconds allowed for a disk to spin and stabilize, that means I/O could cease entirely for almost as long as a minute, which is unacceptable behaviour in a filer.

Really? We are not talking of dimensioning a bridge or safety question. So load balancing is not engineering discipline? When an engineer has to build a toy factory he should dimension it regarding just November/December production request even if the rest of the year, demand is 10 times lower? Not average dimensioned and production balancing over months? All engineers that design nowadays laptops are wrong because cooling system is made to be effective on TDP values, while CPUs overlap this value during boost time. Intel himself is wrong using inertia of cooling system for this turboboost because during this time it overtakes cooling capabilities?
Power supply has to start all drives together (boot up) so IT IS properly dimensioned. We are just talking of improving stability and reducing load on system.
No need of suppressing I/Os, no need to create any dependency between drives, just a delay before sending ATA commands. in case of long idle. It already exists something similar called NCQ (but instead of re-ordering, you would just add tempo). And no need to exaggerate, I gave an example, and because your power supply has to start all drives together, you most probably can group your hard drives in 2 or 4 groups. So even with a huge delay of 5 sec, as you suggest, we are talking of tens seconds.
YOU find unacceptable, and your statement, with some other contributors, is that what you find unacceptable is unacceptable for everyone. So whatever.