More like this



The Channel

PCM prototype beats PCIe flash

Student boffins' box slower at large writes though

The initial performance is impressive; Moneta can read large amounts of data at up to 1.1GB/sec and write it at up to 371MB/sec. For smaller accesses of 512 byte blocks it can read at 327MB/sec and write at 91MB/sec. The team claims this is "between two and seven times faster than a state-of-the-art, flash-based SSD". The 1.1GB/sec speed is half that of the emulated hardware prototype.

Onyx cards installed in Moneta enclosure

Onyx PCM cards installed in Moneta enclosure (UCSD)

Most PCIe flash product uses IOPS as its small block read performance measure. OCZ's VeloDrive is different and that does 29MB/sec with HW RAID when reading 4K blocks and 58MB/sec when writing them. Moneta is more than ten times faster at reading, but only about half as fast as the OCZ card when writing. For a student boffin's team prototype, that's quite amazing.

With software RAID the VeloDrive's sequential read bandwidth tops out at 1.05GB/sec, about the same as Moneta. It can write sequential data at up to 1GB/sec as well, which is more than 2.5X Moneta's maximum write speed. Micron's latest P320h PCIe card writes sequential data at 2GB/sec and reads it at 3GB/sec. Back to the drawing board, Moneta boffins.

PCM software

The Moneta team had this to say about legacy storage controlling software based on hard disk drive storage:

This legacy takes the form of numerous hardware and software design decisions that assume that storage is slow. The hardware interfaces that connect individual disks to computer systems are sluggish (300 MB/sec for SATA II and SAS, 600 MB/sec for SATA 600) and connect to the slower “south bridge” portion of the CPU chip set. RAID controllers connect via high-bandwidth PCIe, but the low-performance, general-purpose microprocessors they use to schedule IO requests limit their throughput and add latency.

Software also limits IO performance. Overheads in the operating system’s IO stack are large enough that, for solid-state storage technologies, they can exceed the hardware access time. Since it takes 20,000 instructions to issue and complete a 4 KB IO request under standard Linux, the computational overhead of performing hundreds of thousands of IO requests per second can limit both IO and application performance.

We show that a redesigned IO stack combined with an optimised hardware/software interface reduces IO latency by nearly 2× and increases bandwidth by up to 18×.

Next steps

The team hopes to build a second generation Moneta device in the next six to nine months, which implies that Micron will be bringing out second generation PCM chips. We have asked Micron about this but haven't had anything back yet.

Steven Swanson, professor of Computer Science and Engineering and director of the Non-Volatile Systems Lab (NVSL) at UCSD, foresees commercial products within a few years; so we are looking beyond 2014 before actual PCM devices appear. He says work will need to be done on the software front to fully realise the benefits of the PCM hardware.

Such software advances could benefit NAND flash devices as well. Based on the boffins' box, PCM is still promising but is far from grinding NAND performance into the dust, and is slower at sequential data IO. Also, it my well be that HP brings out Memristor storage devices before then.

The boys at UCSD have more to do. However, since Micron's PCM people will be only too aware of what the P320h can do, they will be well primed up with the bar that any second generation PCM chip has to get over.

The Moneta system will be shown at the Design Automation Conference (DAC 2011) in San Diego, 7-9 June. ®