Hacker News Reader

Comment by derefr | Hacker News Reader

Hah, I misinterpreted "storage" here, and thought this was going to be about Li-ion cells getting cheap enough to replace the capacitors in the Power Loss Protection circuits in hard drives/SSDs/NVMe boards.

Which doesn't make sense, after giving it a second's thought—batteries wear out after just a few years, while capacitors (very nearly) don't, so it'd put a hard lifetime on drives. (Though, for server use-cases with constant workloads, they have hard lifetimes anyway...)

But it's an interesting question! If—rather than UPSes as specialized enterprise-level hardware—we instead had Li-ion cells on computer motherboards (picture a big brother to the CR-2032 CMOS battery) that kept the disks alive for up to, say, 30 minutes after power loss—could we architect our storage subsystems differently?

(Sorry if that's a complete tangent from the topic at hand, but I'm not sure where I'd post this otherwise!)

View on HN

Keeping the disks alive is certainly doable. At a rough estimate of 10W per drive, even storage servers would only need very moderate batteries. But what would you gain by that?

More interesting (in my eyes) would be to power down spinning disks and processors, but keep caches and ram alive. Then once power is restored, you could basically supply power and continue where you left off, without any reboots etc.

Then again, power loss should be fairly infrequent, and you have to architect against connection cuts or physical disruptions anyway.

Congratulations, you just invented the ACPI S3 power state. Also known as standby or suspend to ram. I've been pretty happy with it since ~1998 when I started returning hardware that didn't support it properly. Although these days even linux on a laptop tends to work properly with S3 and S4 (hibernate, or suspend to disk). That said the concept was also part of the earlier APM standard which was a lot more hit/miss.

Anyway, its great, I used it today to ride out a nearly hour long power outage with a couple of fairly trivial UPSs connected to my NAS, desktop, etc. The desktop was already in standby because that is its default state after being idle for 30 mins. But the NAS went into that mode when the UPS indicated 70% battery life.

The problem of course is that UPSs/etc have miserable power conversion efficiencies when supplying just a couple watts of power, like 10-30% efficiency, so they tend to have max run-times less than 2 hours even when completely unloaded.

More on point, for servers, if your server is lightly loaded, or loaded for only parts of the day, using WOL to wake it and going into standby after a few minutes can save a fair amount of power. With SSDs, the resume times are generally less than a second unless you have bad cards/drivers that take forever to reinit.

If you have a look through some of the Google datacenter tours you'll see they actually implement this idea on a massive scale, though I believe they use lead-acid batteries... Every motherboard in the datacenter has a small battery capable of powering it for a long enough period of time for the backup generators to kick on and take over.

It's definitely a clever idea! Makes maintenance much less risky as it's distributed, and also reduces the chance of having a single large UPS fail when it takes over the load!

I guess the hardware does exist, but I was thinking of a different use-case for it: not for fault-tolerance per se (i.e. waiting for the power to come back on or the generator to kick in), but rather for giving the disk cache up to 30 minutes (after the rest of the machine is well-and-truly cut from power) to flush to the disk itself before finally shutting off, rather than a few seconds to empty its pending writes and park the disk head.

You could, therefore, have a really, really big disk cache, and a relatively slow-to-flush disk. With such an architecture, you could build systems that use disk+cache the way Optane NVMe is being used, without having even needed to invent Flash memory to get fast, highly-parallel writes first.

This is how battery-backed RAID cards used to work. An on-card rechargeable battery could keep the write buffers alive for about 24-48 hrs to flush to disk after you restored power in your datacenter.

Now, they use super-capacitors and a flash memory device to flush the volatile RAM buffers to flash memory, then on next power up the firmware can flush the flash buffers to disk. This all happens within the RAID controller so the host OS just thinks it already got write confirmation back when the PCIe transactions finished and the data blocks landed in the controller's RAM.

> kept the disks alive for up to, say, 30 minutes after power loss

Not relevant to disks, but you could have battery-backed DRAM? Commonly used in RAID clusters as a cache.

The main problem with storage architecture is too much hardware that lies about whether it's actually committed a write.

Portable computers sort of do that.

I had a vague idea that I had heard something similar before:

http://www.zdnet.com/article/google-reveals-secret-server-ha...

Reply on news.ycombinator.com