MORE IOPS
After seeing how many 512 byte IOPS the P320h was generating in the SNIA IOPS loop, we conducted a little experiment. The P320h is clearly able to generate more IOPS than we saw with 100% read 512 byte IOPS, but we were limited by CPU cycles. The performance we experienced was CPU limited because it requires tons of CPU to create a workload that can exercise such a beastly drive. We set our testbed’s processor to 3200MHz, 4000MHz, and 4900MHz to see what would happen. How many IOPS could we generate?
As it turns out, 1.74 Million 512 byte IOPS was the best that could be done with a retail SKU i7 3930K at 4900MHz. That’s unbelievable. One million is already as many as four Intel 910s can generate, but the P320h hits the afterburners and keeps on going, getting an additional 70%. After discussing this with Micron, they reckon that 1.9 million could be the limit for 512 byte IOPS. That would still only amount to 1000MB/s of throughput, but it’s still supreme among all drives we’re aware of. With some more horsepower under the hood, we could certainly achieve slightly more.
REPORT ANALYSIS AND CONCLUSION
This evaluation focused on Linux performance, but we did run tests under Windows as well. Windows and Linux both schedule IO a little differently, and Redmonds OS has some additional overhead that can prevent the P320h from hitting quite the same heights as in Linux, but nothing too significant. Its mainly a few percent here and there, but Linux performance is better, if only by the slimmest of margins. Given the amount of CPU it takes to even put a test load on the drive with our bench, the more efficient OS will result in higher performance in our testing.
The P320h has been put through a brutal testing regime for hundreds of hours. It has not been found wanting, either. Its as-billed by Micron, with some surprising twists and turns to keep things interesting.
Its true that the P320h was first announced almost 20 months ago, but it looks as though the results were worth it. The custom ASIC will undoubtedly be utilized by future products, and unlike some other high end PCIe solutions, it doesnt require any host resources to perform any of the flash processing. All management functions are performed by the controller, saving valuable host resources for other endeavors.
The new Micron flash processor might even end up in client solutions one day. For more modest applications, the controller can probably be modified into 8 and/or 16 channel versions, depending on how big of a hammer is needed. Despite the beefy brains and lots of flash, the whole drive fits within a 25w envelope (though you can allow it to use a few more watts for writes if youre confident in the hosts motherboard specs). It would be hard to say how many hard disks would be necessary to achieve three-quarters of a million IOs per second, but at 200 to 500 IOPS per spindle, it’s a whole lot.
Read performance is out of this world, and write performance isn’t shabby either. Those 785K/205K+ 4K IOPS numbers Micron appends to the P320h’s datasheet are all day, every day numbers. With a little extra over-provisioning, we can even get those 4K write IOPS significantly higher. Also, getting 1.7 million 512 byte read IOPS is always spectacular, even if few applications demand that kind of performance. At every block size, and read/write mix, the P320h excels.
It remains clearly more suited for read-centric workloads, but it does manage to hold its own on writes, too. Out of the box, it puts down nearly 500K 4K write IOPS, but quickly settles into steady state when writing to the entire LBA space.
The P320h HHHL is a flagship product, one which highlights the advantages of vertical integration.
Owing the controller, fabs, firmware, drivers, and DDR3 cache has its benefits, and the number of companies able to match an all-proprietary drive are few. Considering the humble origins of Microns first solid state drives, the progression to their first PCIe SSD has been rapid.
Micron may or may not ever release a custom controller for the client space, but its good to see the company go their own way for the pure enterprise P320h.
Micron doesnt own the controller. It is made by IDT.
We are aware of that, thanks. Our reasoning behind wording as such is because this is, by no means, a simple stock implementation of a controller and similar could not have been accomplished without Micron’s engineering expertise and software. Great point and perhaps we could reword things just a bit…
Micron has a Minneapolis-based controller team which did much of the work on the controller. Basically, IDT has a stock PCIe controller, but it’s easily modified for custom jobs. Micron refined the design for the P320h. IDT now has a reference NVMe design, but the NVMe standard is far from universal yet. One day, a PCIe SSD won’t need a special driver, but today they do.
Micron developed and owns the chip, IDT just fabs it.
Incorrect. This is the very same controller that is used with the new NVMe controllers that IDT has developed.
Just to help you out, this is what has been posted at Anands after they inadvertently stated it was NVMe.:
Update: Micron tells us that the P320h doesn’t
support NVMe, we are digging to understand how Micron’s controller
differs from the NVMe IDT controller with a similar part number.
Our interpretation of the chip appears to be correct as it is written and this same ‘structure’ has been used in the SSD industry prior. This is not a simple plug and play adaption of a chip, but rather, custom package.
Thanks again.
Yes, it isnt NVMe, but it is an IDT chip, therefore it is not developed in house by Micron.
old news
Just needs a few heat sinks and a fan or maybe a water block to keep it cooler.
Todd – What makes you think you know so much about this chip?
Is the RAIN implementation safe enough to use without RAID 1 running outside of it (say across 2 350GB cards) it sounds good, but if you have a firmware or controller related failure you’re still at risk right?
Is this bootable? And just for kicks, what would the as-ssd results be?