Author Topic: OCZ's HSDL: A New Storage Link For Super-Fast SSDs  (Read 1584 times)

Offline javajolt

  • Administrator
  • Hero Member
  • *****
  • Posts: 36193
  • Gender: Male
  • I Do Windows
    • windows10newsinfo.com
OCZ's HSDL: A New Storage Link For Super-Fast SSDs
« on: October 01, 2010, 07:13:09 AM »

SATA 6Gb/s not fast enough? OCZ's new High-Speed Data Link serves up to 2 GB/s through a PCIe-like connection. Complementing the interface technology is a similarly-new 3.5" SSD called IBIS. Rated for up to 120,000 IOPS, enthusiasts should take note.

SATA 3 Gb/s treated us well for a very long time. Even today, mechanical hard drives simply cannot push enough data to saturate a 3 Gb/s link. Shoot—most SSDs can’t saturate a 3 Gb/s link, either. And now we have 6 Gb/s connectivity being included on AMD’s SB850 southbridge. In the next couple of months, it'll also surface on Intel’s P67 chipset. On the desktop, though, these fast storage interfaces just aren't super-exciting.

In the enterprise world, 6 Gb/s signaling is already being exploited. The faster transfer link between a compatible RAID controller and midplane means you can attach twice as many drives to each port before running out of headroom. It’s a simple matter of math, really. If one 6 Gb/s link is capable of up to 600 MB/s, assume you can attach three or four disks to an expander and not see a performance bottleneck, even if they’re all running full-out. In the days of 3 Gb/s connectivity, it would have been possible to force a limitation much sooner.


The controller board, one of three PCBs inside OCZ's IBIS

Of course, the desktop market isn’t really accustomed to seeing SAS expanders or SATA port multipliers. In the world of the enthusiast, you’re usually looking at one device per port. And at 6 Gb/s, we’re now a long ways away from seeing a single drive saturate a single link.

Why, then, is OCZ announcing its own interface, labeled as the answer to current interconnect limitations? That’s the exact question I posed to Ryan Petersen, the company’s CEO.

OCZ Announces Its High Speed Data Link

New technologies generally don’t gun for the mainstream. Rather, they shoot for the top and derive their way down into more affordable versions that enthusiasts with car payments and mortgages can justify. Such is the case here. A high-speed link suggests the need for more sequential and random I/O throughput than what’s available today through SAS. And so Ryan says he’s going for the ultra-high-performance market—anything from supercomputing to bandwidth-starved (and presumably fairly wealthy) power users.


One of two storage boards inside our 240 GB sample

To a degree, anyone in the range of flush enthusiasts to Oracle database administrators probably doesn’t care what bleeding-edge storage technology costs. The thing is, OCZ isn’t trying to translate its High-Speed Data Link into some new class of inaccessibly-expensive storage products. In fact, a first-glance at the company’s new SSDs based on HSDL shows a handful of price points that actually make sense.

Alright, so $530 for 100 GB of storage space is a little ridiculous. But you’d expect to pay $800 for 320 GB of X25-Ms in RAID 0, so $1100 for a 360 GB IBIS drive isn’t outside the realm of reason if the performance is right. And if you were specifically shopping for something that could keep up with HSDL from the SAS world, you’d need a RAID controller, which would drive up cost. OCZ bundles each of these drives with the controller card needed to make an HSDL connection.

Interestingly, the IBIS drives are more expensive than OCZ’s recently-released PCI Express-based RevoDrive. The company’s official stance is that the RevoDrive is aimed at optimized price/performance, appealing to enthusiasts. The IBIS is better suited to workstations and the HPC market. Backing these projections are OCZ’s initial performance numbers. The company is claiming read and write speeds of up to 740 MB/s, and up to 120,000 IOPS. Those figures are insane. Clearly, they’re also in excess of what a single SAS or SATA interface could accommodate. And according to OCZ, this is only the first generation of the technology—it’s supposed to get faster from here.

Intrigued? I was too. Let’s look at the HSDL interface itself for more on what it purportedly offers.

OCZ isn’t doing anything particularly tricky with its HSDL interface. Everything happening under the hood is very standardized—and OCZ is very forthcoming in describing the foundation on which HSDL is based. HSDL employs the same 8b/10b encoding scheme as PCI Express, SATA, HyperTransport, and USB 3.0. And once PCI Express 3.0 catches on, HSDL will even accommodate the 128b/130b encoding that interface specifies.

For the most part, PCI Express is the center of HSDL’s world.

A single HSDL channel consists of four PCIe lanes bonded together, yielding predictable scaling. Right now, the company is offering its IBIS SSDs with single-channel adapter cards that (obviously) drop into a x4 motherboard slot. In the future, it plans to scale up to dual-channel cards that’ll drop into a PCIe x8 slot.


The math is easy. In a second-gen PCI Express x4 slot, HSDL serves up 2000 MB/s of throughput. In a x8 slot, a dual-channel card would push up to 4000 MB/s. Of course, the dual-channel card OCZ plans will also sport four ports, limiting cumulative throughput to roughly 1 GB/s per drive (still more bandwidth than a single first-gen drive can saturate).

In between the controller and the drive, OCZ is using 4x SFF-8087 connectors. While it might seem a bit ironic to use SAS connectivity to outperform the SAS interface, OCZ says the four-channel connector gives it the number of LVDS pairs needed for HSDL. Each channel consists of four bi-directional pairs, correlating to the four bonded PCIe lanes.


The magic happens at each end of the cabled connection, where a logic chip serves as a signal driver and buffer. In the case of a dual-channel controller attached to a single drive, that piece of logic also multiplexes to create a single ultra-wide “PCIe x8” link. If you want to get a little more technical, OCZ is using the logical PCIe layer over the company’s own embedded physical layer. On the drive itself, you have a conversation chip that turns the SATA into a PCIe-type signal (Petersen specifies that it isn’t actually PCIe), which is serialized and re-buffered before getting sent out.

On the controller side, another chip serves as the basic physical protocol conversion to turn the physical layer into PCIe. Should a motherboard vendor wish to integrate HSDL support, it’d need the SAS connector and the conversion chip found on OCZ’s controller card—everything else is drive-focused. “The goal,” according to Petersen, “is to let motherboard vendors be able to implement HSDL at a five-dollar cost. If they can’t do it for less than $5, it’s a bad move for us to have the product.

Using this architecture, OCZ expects to see 500 MB/s per LVDS pair ramping up in the future. Of course, even with this first-generation implementation of the technology, HSDL is pretty much overkill—the performance of today’s controllers and NAND flash isn’t high enough to saturate a four-pair, 2 GB/s single-channel HSDL interface. OCZ’s Petersen teases that this will be addressed soon. With that said, the company expects today’s single-channel controller to handle the performance of first- and second-gen IBIS-style SSDs.

Three generations down the road, more throughput should start to become necessary (at least, that’s the game plan). When that happens, the company can start using standalone PCIe 3.0 transceivers and encoding chips (which are actually available right now) to adopt 128b/130b encoding, increasing throughput even more. There’s simply no need for that extra headroom today, given the available drive technology.

The IBIS SSD


But I couldn’t resist dismantling the thing.

Inside the IBIS, you’ll find three boards. There’s one controller board hosting the SAS/power connections and a 3 Gb/s Silicon Image SiI3124 PCI-X-to-four-port-SATA controller, which snaps into an internal backplane. Two other boards are loaded down with 34 nm MLC NAND from Intel, along with a pair of SandForce SF-1200 controllers each. Each one of these appears as a 55 GB drive in the Silicon Image controller BIOS.


Silicon Image four-port SATA 3Gb/s controller

If it helps, think of the IBIS as very similar to the RevoDrive X2 that SandForce was showing off at IDF 2010. Only, instead of dropping into a PCI Express x4 slot, the IBIS plugs into a controller that drops into a PCI Express x4 and communicates over an interface that is in no way throughput-constrained today. Architecturally, though, both drive products are really very similar.


BIOS


Of course, the IBIS is bootable, and although Windows 7 doesn't include the driver for Silicon Image's controller, adding once Windows 7 setup reports that it didn't find any drives lets the OS see the SSD and install to it. Although there are 256 GB worth of NAND chips and OCZ advertises our sample as a 240 GB offering, you actually get 223 GB of usable space. This is a result of 1) a binary to decimal conversation and 2) over-provisioning on OCZ's part.

Sandforce


SandForce SF-1200 controller (x4)

The sample OCZ sent over boasts 128 GB of flash per board (256 GB total), though the drive’s advertised capacity is 240 GB. In essence, what you’re getting is an array of 128 GB flash-laden boards inside that 240 GB model. Technically, this isn’t actually RAID 0; OCZ employs Frame Information Structure (FIS)-based switching, allowing the Silicon Image IC to talk to all four SandForce controllers simultaneously and facilitating the aggregated bandwidth you’d normally chalk up to striping. Presumably, the lower-capacity drives don’t employ as many channels per board, resulting in performance figures that trail the 240-960 GB SKUs. OCZ’s data sheet claims maximum sequential reads of up to 740 MB/s and writes as high as 690 MB/s, with up to 100,000 IOPS for the 100 and 160 GB drives.

OCZ does have plans to sell a separate RAID card that’ll allow you stripe physical drives together—the board needs another logic chip to mux/demux the drives’ channels. The company knows that there are certain features that’ll need to be enabled if it hopes to see widespread adoption, and it says they’re being worked on currently. RAID support is one of those features. S.M.A.R.T. compatibility is another.

The IBIS drives don’t support TRIM in any form today. OCZ claims that this could potentially change in the future, but assume it won’t if you’re shopping around today. OCZ also makes no mention of idle garbage collection support, though past experience suggests this shouldn’t be a serious issue on a SandForce-based product. Should we experience performance degradation after prolonged use, we’ll make an update to this piece. Update: I followed up with OCZ, and it sounds like the SandForce controllers perform their own form of garbage collection, wrapped up into the DuraClass technology package. This should help maintain the drive's performance.


One side of one storage board inside the IBIS

Also, be sure you’re dropping the IBIS into a system with ample system throughput. Naturally, using the drive on a P55-based platform already rocking a discrete graphics card is going to kill its performance (thank Intel’s chipset folks for that). You’d almost assuredly need an X58- or 890FX-based desktop platform, or something based on 5520 in the workstation space. Then again, if you’re in the market for a drive like this, “going cheap” probably wasn’t in the cards anyway.

The Future Of IBIS

In discussing IBIS, OCZ’s Ryan Petersen gave away some of the company’s plans for the drive’s future (it does seem to have a roadmap—it’s not a one-off product). While the architecture being used today is capable of up to 120,000 IOPS, OCZ expects to release a version in three months capable of twice that. From there, this first-gen HSDL implementation will be maxed out, and the interface speed will have to increase. Once they’re there, today’s hardware is going to get pushed down-market (a la Nvidia’s G92, for example), and a new generation will be launched up top.

Test Setup And Benchmarks


Benchmark Results: PCMark Vantage






Although the IBIS drive bests a single X25-M and a pair in RAID 0 every step of the way, it's the Media Center Performance metric that demonstrates just how substantially the PCIe-connected drive can smoke conventional SATA-attached SSDs.


Though we don't see the maximum performance ratings cited by OCZ in this close-to-real-world collection of tests, it's clear that there's a substantial performance benefit to be had from IBIS and the HSDL interface it employs.

Benchmark Results: IOmeter, Test Patterns


This is the sort of chart that a database admin loves to see. Our workload include queue depths of 1, 2, 4, 8, 16, 32, and 64, and performance increases each step of the way, culminating in nearly 55 000 IOPS. The average of 43 554 is still incredibly impressive, though, easily besting the Fusion-io ioDrive we reviewed a while back.


Benchmark Results: Streaming And 4K Reads/Writes


Our streaming reads benchmark pattern actually shows the IBIS drive pushing more than OCZ's rated 740 MB/s. This result is on par with the ioDrive and ioXtreme Pro products we've tested.


Writes fall off a bit, as expected, but not nearly to the extent of Intel's X25-M--even in RAID 0. Again, these results are close to some of the fastest PCIe-based SSDs we've tested.


Although we've seen 4k read performance similar to OCZ's from product like the ioDrive, the 4k write performance is unprecedented. We already knew Intel's X25-M would fall off precipitously here, but this is the first time we've seen 100 000+ IOPS in this benchmark pattern.

Conclusion

OCZ's HSDL Interface

OCZ's IBIS is an interesting piece of hardware, but I'd argue that it isn't the focus here today. Rather it's the company's High-Speed Data Link in the hot seat, theoretically able to move up to 2 GB/s over a very PCI Express-like bus.

Unfortunately (or fortunately, depending on how you look at it), even a 3.5" drive with four SandForce SF-1200 controllers and 240 GB of Intel NAND flash cannot come close to saturating the interface. Without question, we'll have to wait for the mysterious next-generation IBIS drive OCZ's CEO Ryan Petersen mentioned.


With that said, the stage is set for an interesting evolution. A single HSDL channel and a first-generation drive outpaces what one SAS connection can do, throughput-wise. And OCZ is talking about multiplexing lanes to generate even more performance. You can't do that with SAS. SAS is great for taking four ports, hooking them up to a midplane, and attaching 20 drives. But if performance is your one and only care in the world, you can't take those same four SAS links and route them into a single drive. Even if you could, we haven't seen a single SAS-based device able to push the interface's maximum 600 MB/s--much less the numbers OCZ is throwing around.

It might seem silly, then, to already be talking about a two-channel PCIe x8 card able to move 4 GB/s. Armed with four ports, though, it's almost mind-boggling to think about what a quartet of IBIS drives could do for highly transactional workloads in a single server. I'm almost positive that it'd be impossible for me to justify spending more than a grand on 360 GB of capacity, as an enthusiast. But I know for a fact that there are plenty of business applications for this--environments where the capacity and cost aren't issues, but performance is.

The IBIS SSD

Alright, so we know that HSDL in its current form theoretically has gobs of headroom to spare. Where does that leave IBIS, the first drive family designed to take advantage of the HSDL interface? The way it stands right now, most of the IBIS drives are, simply, expensive. This isn't OCZ's play on performance for the price--that remains the RevoDrive, according to the company. Here, you're paying a premium for product that can deliver more than 100,000 IOPS and streaming read/write performance in excess of what a SAS port serves up. OCZ says enthusiasts are in its reticle here, but I don't really see any practical reason for power users to go this route, unless you have an application you know is outright I/O-limited. The more receptive markets are going to be Web servers, file servers, databases, and email.

Shoot, if OCZ could partner up with a storage software vendor and create a caching package that'd operate similar to Adaptec's MaxIQ or Intel's SSD Cache premium feature key, it'd have a massively powerful solution for speeding up application performance--like a miniature GridIron TurboCharger.

Bearing in mind that the IBIS seems tailor-fit for the enterprise space, my main reservation right now would probably be a lack of features you normally find in more business-class storage solutions, such as battery backup to prevent against loss in the event of power failure. OCZ's SF-1500-driven Vertex 2 Pro employs a capacitor that stands in as a temporary backup, allowing the SSD to finish outstanding operations before losing power. You don't get that concession on an SF-1200-based implementation like IBIS. That'll be something IT decision makers, who are a notoriously conservative bunch, have to consider before jumping on this technology. We're also missing TRIM support here, and it remains to be seen whether this affects performance over the long term.

Regardless, IBIS is OCZ's first generation of what will hopefully turn into a wider range of HSDL-based SSDs. It's incredibly fast today, and the company is talking about forthcoming versions able to tax its interface in an even more significant way. PCIe-based SSDs are cool, but I'm looking forward to seeing HSDL make ultra high-performance drives more flexible. And that's really what this technology is all about. Four super-fast drives plugged into an HSDL PCI Express card is much better in a server than four drives plugging up their own slots.

source:tomshardware