Storage Devices at WinHEC 2008

wdkblog — Thu, 27 Nov 2008 08:40:00 GMT

Quite honestly, spending most of my time deep in the innards of the kernel file system stack, I wasn't entirely convinced that WinHEC would have a lot of relevant information to offer - boy, was I wrong!

As a matter of fact, WinHEC provided so much info that, for the sake of brevity, I am "forced" to focus on just one aspect that I found particularly interesting: storage devices.

From the WinHEC presentations I attended, "serious" storage devices essentially come in three flavors:

· Solid state disks (SSDs)

· Traditional hard drives or hard disk drives (HDDs)

· Hybrid drives (SSD + HDD)

And each flavor has its pros and cons:

Solid State Disks (SSDs)

Probably the SSD that most folks are familiar with is your garden variety flash thumb drive. You pop it into a USB port and voilà, you have a new "disk drive". Well, let's just say that there are prototype SSDs that make a thumb drive look like a lethargic gnat. The prototype I saw, implemented on a PCIe card, was very performant. However, despite their performance, every silver lining has a cloud:

Pro	Con
Speed: Prototype SSDs have impressive IOPS (I/O operations per second), particularly for reads. The PCIe prototype I saw spent much of its time waiting for the rest of the system to respond.	Endurance: Currently, SSD NAND flash chips tend to wear out with use. In general, you need to erase NANDs before writing to them (write-in-place is not an option as it is for rotating drives) - you're good for about 100,000 erases for SLC (single-level cell) flash and about 10,000 erases for MLC (multiple-level flash). Ultimately, as the number of erases goes up, so does the number of device errors, which is why wear-leveling (typically implemented in the firmware of the device) is critical for SSDs. (Wear-leveling attempts to evenly spread the erases around on the device, increasing device endurance.)
$/IOPS: For a typical consumer, dollars per I/O per second ($/IOPS) really isn't that important - low $/GB is what's critical. However, for performant environments, $/IOPS can trump $/GB.	$/GB: SDDs currently have a relatively high dollar per GB value as compared to HDDs.

Interesting Tidbits

· In order to make intelligent wear-leveling and NAND block cleaning decisions, file type metadata can be important. For example, photos, music, and videos tend to be "cold" read-only-ish data. Knowing what's "cold" and what's not could be very important in choosing which blocks to write data to, from a wear-leveling perspective.

· SSDs can function as either slow memory or a fast "disk" - they've created a new tier in the system. Likewise, both memory teams and/or traditional disk teams at various companies can develop SSDs.

· SSDs are starting to encroach in the optical media space - flash is now large and cheap enough to do so.

· Typically, if you want more performance from a computer, you buy more memory and a bigger hard drive. One study showed that you can get the same amount of performance with a smaller amount of memory plus an SSD (for search, database, and similar applications). So for applicable applications, an SSD can reduce operating costs (despite their endurance issues).

· Currently, there are not international metrics for SSDs. For example, how do I choose a device with high endurance? What is the current "health" of my installed SSD device? Ratings for important SSD attributes need to be hammered out.

· Can Windows boot time and file system indexing be increased via an SSD? MS is working on these questions now.

Hard Disk Drives (HDDs)

HDDs are, of course, your traditional rotating magnetized-disk-o-metal with a swing arm and various other mechanical components. Despite their "old school" nature, HDDs are undoutably not going away anytime soon. Historically, there have been many contenders but HDDs have always survived and excelled. Generally, it's $/GB that drives storage device sales, and given today's widespread use of digital cameras and camcorders, the need for cheap storage is only going to grow. One study predicts that 2011's storage needs will be about 10 times that of 2006's.

Pro	Con
$/GB: So far, HDDs provide the biggest bang for the bug with respect to this metric.	$/IOPS: Compared to SSDs, HDDs have relatively high $/IOPS.
Endurance: Compared to current SDDs, HDDs have outstanding endurance characteristics.	Slow: Compared to SSDs, HDDs can seem downright sluggish (particularly against SSD read operations).
Extensible: HDDs can easily be clustered, placed in RAID arrays, etc.

Interesting Tidbits

· One WinHEC presenter felt that SSDs currently don't have sufficient capacity - they'll need about 10x their current capacity to compete with HDDs.

Hybrids (SSD + HDD)

Hybrid drives have been around awhile in the sense of "small" memory caches on HDDs to help speed up access to frequently accessed data. The hybrid device, as described at WinHEC, is much more than this - it would contain a significant SSD component as well as a hefty HDD component.

To me, hybrid drives really offer the best of both worlds. For example, the hybrid firmware could keep an eye on the health of the SSD component (i.e., notices when NAND chips are starting to fail) and could guarantee reliability by shunting data to the HDD component as appropriate (at the expensive of device IOPS, of course). One presenter's contention was that successful storage device manufactures will need to provide all three flavors classes: SSD, HDD, and hybrid.

Call to Action

If you've ever attended WinHEC, I'm sure you noticed the myriad call-to-action slides at the end of nearly every presentation. Yes, these can be a tad annoying but there's one I would like to pass along which affects file system driver developers. Per the Clustered Shared Volumes folks, please test your file system drivers with the CSVFilter.sys minifilter in the stack (i.e., any data transformation drivers found below CSVFilter.sys will result in all I/O to remote volumes being done over SMB, with CSV Direct I/O not being used for that particular CSV volume - this can significantly impact performance).

Finito

Well, that about wraps things up for me - granted, it's the 10,000 foot WinHEC overview of storage devices (at least my interpretations thereof). If you'd like less than a 10K' overview, definitely check out WinHEC's 2008 Conference Presentations - there's a lot of stuff here - the relevant presentations for this article were as follows:

1. Design Tradeoffs for Solid-State Disk Performance

2. Solid-State Storage in Server and Data Center Environments (should be available soon)

3. New Developments in the Storage Platform

4. Windows 7 Enhancements for Solid-State Drives

5. Is Your Disk Drive Going Away

6. Enhanced Storage Support in Windows 7

7. Windows Virtualization and Clustered Shared Volumes

Hope you've had a bit of fun from this post, take care!

- Karlito Bonnevie [MSFT], Programming Technical Writer, Installable File System Documentation

Windows Driver Kit (WDK) Documentation Blog : Storage

Storage Devices at WinHEC 2008