Micron Demos PCIe RealSSD P320h, Achieves Over 3GB/s Of Sustained Throughput | Hacker News Reader

Micron Demos PCIe RealSSD P320h, Achieves Over 3GB/s Of Sustained Throughput

hothardware.com

34 pointsbigwophh19 comments

Fuck yes.

This isn't about GB/$.

This is about IOPS/($ x density) and Throughput/($ x density). Databases and datastores used to be limited by iops, and SSD's are altering the balance of where that bottleneck occurs: we've already seen absurd 9x SSD systems[1] and 24x SSD systems[2] running off three or four SATA/SCSI cards; this offers those levels of performance in a single package.

3GBps is 24Gbps. A mere two of these will max out the IO Hub on a Nehalem, which is good for 40Gbps[3]. As for getting that data into the CPU, the top of the line MP Xeon offers four QPI links good for 25.6Gbps a piece. That's 100Gbps, which four of these could provide, and it stands as a fixed upper bound for how much data an SMP (4-way in this case) system can be fed (and thus crunch) at once (and uses expensive 4-way cpus to do it). The throughput limit is now (erm, will be) the platform, not the IO storage. This is huge. Wait, no, strike that, it's incomprehensibly small, small and fast.

References:

[1] "Battleship mtron" http://www.nextlevelhardware.com/storage/battleship/

[2] "Samsung SSD Awesomeness" http://www.youtube.com/watch?v=96dWOEa4Djs&fmt=22

[3] "Building a Single Box 100Gbps Router" via http://shader.kaist.edu/packetshader/

I'd love to hear about the ASIC this runs on.

Will it present SATA's AHCI interface? Will it eschew SATA altogether?

AHCI is a serious bottleneck because it requires PIO reads and has a queue depth limit of 32. PCIe SSDs should use the NVM Express interface, although this card is probably too old to support it. http://www.intel.com/standards/nvmhci/index.htm

AHCI supports DMA, which has been the preferred transfer standard in ATA since the mid-1990s. Nobody in their right mind would intentionally use the PIO modes except on very very old hardware.

And for a device that can handle 320,000+ IOPS, a short queue depth won't be a bottleneck.

Sorry, I meant that AHCI requires the driver to do a PIO read to set up each command, even though the data is transferred using DMA.

This card has 32 flash channels and 128 or more NAND dies, so a queue depth of 32 would leave most of the dies idle.

I'm not sure it's relevant. Queueing in the transfer protocol is a useful feature to address the limitations of physical hard disks (i.e., to delegate the ordering of temporally-proximate I/O operations to the device, which usually can do it better than the host OS can). With flash disks, optimization of seeks is unnecessary (since there are no seeks), and they are usually so fast that the command queues rarely grow.

What's the crash-safety/power-loss durability on these guys?

The press release mentions OLTP workloads, but if they don't have supercaps or a battery or something, you'd have to be an absolute idiot to trust your data to them, no matter how compelling the performance numbers. The pictures don't show any connectors for a battery or capacitor-looking doodads (or at least anything I recognize as such), so I'm extremely skeptical. Unf., the linked site is down for "maintenance", and some quick googling yields nothing on-point.

(Edit: formatting)

If they don't cache writes, and they honor fsync calls, then not having a supercap isn't such a liability. That said, when a vendor announces a enterprise-targeted storage product without mentioning reliability, it likely isn't there.

There's never any excuse for failing to perform your own power-plug tests.

Would be amazing to see one of these built for Thunderbolt since it is essentially PCIe over a serial interface.

We're at a point where regular hard drives are ridiculously cheap $60-80 for 2TB (2.9c - 3.9c per GB), while regular SSDs are $1.5-$2 per GB.

There's a huge gap in between.

Hybrid SSD/HDD drives were supposed to fill that niche, but they are still pretty crappy, they don't really solve any HDD problems.

One of my biggest questions is the lifetime of these drives. I know that flash memory has a certain number of write-cycles but has there been any study done on how long these drives last in an almost constant write environment like OLTP?

IIRC, SLC NAND has ~100,000 write cycles and MLC ~10,000 (enterprise-level SSDs almost exclusively use the former). I'm not aware of any recent studies specifically on durability, but given that there are no moving parts I'd imagine that SSDs last at least as long as mechanical drives. There's also this bit from the press release: "[The] 700GB drive [is] able to write 28 terabytes of data every day for five years."

There's a big tradeoff between reliability and density: 50nm MLC is good for around 10k cycles, but 34nm and 25nm MLC are rated in the 3k-5k range, and it's not clear yet whether 20nm MLC will be able to stay in that range. SLC is much more reliable, but also much less dense than MLC.

It's easy to maintain overall device reliability (and performance) by having more spare area, but that negates a lot of the benefit of moving to a smaller process. (And the performance penalty from having fewer flash chips to stripe across for a given drive size also detracts from the smaller process size.)

Warranty is key here. That and having prepared ops. Which is no different than it was with HD's.

But for the sake of it, let's run the numbers.

A 700GB drive being written at 2GB/s would fully cycle in 350 seconds. That's 86400/350 is 247 cycles a day (amusing 247 ~ 24/7 coincidence spotted).

Assuming a 10,000 cycle life, that's 40 days of operation. Go to 100,000 cycle life and that a 400 day lifetime of non-stop full-speed writing.

And worth every sweet cent of that somewhat short life, in this commenters opinion. I'm sure that the half rack of HD's it'd take to get this kind of IOPs wouldn't have an in any way improved expected lifetime cost. If you just need throughput the story may be a little different.

"The P320h drive also provides enterprise-class levels of endurance with the 700GB drive able to write 28 terabytes of data every day for five years."

Not only that, but the write endurance failure mode is benign: the drive simply becomes read only.

> Hybrid SSD/HDD drives were supposed to fill that niche, but they are still pretty crappy, they don't really solve any HDD problems.

Hybrid drives aim to bring together the (cheap) capacity of hard drives and the speed of SSDs.

And they fail at both.

Hybrid drives won't take off. Just hit the history books. It's not like there were hybrid tape/HDD drives. When a technology like this comes along it will slowly marginalize the old gear.

So my crystal ball predicts the adoption of tiered setups. Anything that needs the performance bad enough to justify the cost will buy SSDs. HDDs will slowly be pushed into secondary "bulk storage" roles, then finally to backup and archival.

It's not like there were hybrid tape/HDD drives.

Actually, for a long time if you requested aerial image data from the USGS, an automated tape library would copy the desired data to hard drive, and then you could download it. It was always hard to find that site (obviously you don't want Google crawling a tape library), and I can't find it now if it still exists.

Granted you may have been referring to actual integrated hardware, but I think such an automated system is close enough.

Reply on news.ycombinator.com