|
Lead
SSDs: the future of enterprise storage
Akhtar Pasha looks at the new approach to enterprise
storage technologyNAND-based flash SSDs in the wake of DRAM-based SSDs
that are still going strong
The
debate for supremacy in disk technologies continues to center around DRAM-based
Solid State Disks or SSDs and NAND-based flash SSDs. While the former have been
used in blade servers and enterprise storage arrays for quite some time, flash
SSDs are non-volatile (they retain data without a power source), have no mechanical
parts and are used in consumer electronic gadgets such as phones, PDAs and MP3
devices.
While HDDs have been the workhorses of enterprise storage for nearly 50 years,
DRAM SSDs and now flash SSDs are getting more attention than ever. There are
important speed differences between these two technologiesflash SSDs are
much faster than spinning disks, but they are not as fast as DRAM-based SSD
when handling I/O. So lets understand how both these two technologies work and
where they differ from one another as well as how these two will be used in
future.
Disk is the bottleneck
|
"Disk
technology has been lagging behind the rest of the computing systems.
While CPU performance has doubled every 18 months, disk technology performance
per gigabyte of storage has been falling. However, SSDs will change the
face of enterprise storage over the next 10 years"
- Surajit Sen
Director-Marketing & Alliances, NetApp
|
|
"Give
NAND SSDs another 5-8 years and it would emerge as the default storage
technology, while DRAM SSDs would remain a special purpose, highly scalable
technology used in mission critical application"
- Ramachandran Narayanaswamy
Vice-President and Head-Networking and Storage Industry Group, MindTree
Ltd
|
|
"Flash
SSDs cannot match the performance of DRAM-based SSDs that are widely used
in OLTP and database intensive environments. DRAM SSDs would typically
operate or give a million IOPS while a 15,000 RPM HDD would give 350-450
IOPS"
- Sivasankaran L
Director-Storage Practice,
Sun Microsystems India
|
To understand this let us examine the disparity that exists
between server processors, memory and disk. Intels latest technology innovations,
including 6-core Intel Xeon 7400 series (formerly code named Dunnington) processors
only highlight a significant gap in the IT ecosystem: multi-socket, multi-core
servers have far outpaced the performance of traditional disk drives resulting
in expensive and complex architectures, which use massive amounts of expensive
DRAM or disk drivesdesigned to maximize CPU utilization.
Dave Hitz, Founder, NetApp in his blog puts it nicely how DRAM-based SSD and
NAND flash-based SSD technologies differ from one another, and what is driving
the innovation in the NAND-based SSDs. According to him, almost every measure
of computer performance increases exponentially with the important exception
of disk drives that keep getting bigger but are not getting much faster. As
a result, the number of seeks-per-second available for each gigabyte of data
(seeks/second/GB) is plummeting. He explained, from a human perspective, that
seeks/second/GB have gone down by a factor of five hundred, but from the CPU
perspective, it is five hundred thousand times slower.
Surajit Sen, Director-Marketing & Alliances, NetApp,
said, The disk technology has been lagging behind the rest of the computing
system such as the server CPU and memory. While CPU performance has doubled
every 18 months, disk technology performance per gigabyte of storage has been
falling. However, SSDs will change the face of enterprise storage over the next
10 years but this technology has its own limitations and one would need to modify
the file system to gain the required level of performance.
Let us take a closer look at some of the limitation of flash-based SSDs vis-à-vis
DRAM-based SSDs and the approach vendors are taking to improve the performance
of flash-based SSDs.
Performance issues
Writes to flash SSDs require that the cell first be erased after
which it can be programmed with new data. The erase procedure (typically ranges
from 1.5-2.0 milliseconds) is quite long compared to everything else in a memory-based
system and nearly as long as a seek operation on a HDD. While SSD Read performance
can be sustained at 20-40 MBps, Write performance significantly lags at only
1-5 MBps.
Ramachandran Narayanaswamy, Vice-President and Head-Networking
and Storage Industry Group, MindTree Ltd., added, In a typical SSD, if
you want to write 4 KB of data, the system first needs to copy a 128 KB block
of data from the flash drive to main memory. Then the system modifies the data,
now in the system DRAM, with the 4 KB of new data. Finally, it needs to write
the entire 128 KB block back to the flash drive. Typical write amplification
multipliers are in the neighborhood of 2040X, on average.
Sivasankaran L, Director-Storage Practice, Sun Microsystems India, said, Ramping
up disk performance in a square-inch is just impossible and there is a clear
mismatch between the speeds at which current server CPUs and memory operates.
Further flash SSDs cannot match the performance of DRAM-based SSDs that are
widely used in OLTP and database intensive environments. DRAM SSDs would typically
operate or give a million IOPS while a 15,000 RPM HDD would give 350-450 IOPS.
Herman Yiu, Regional Marketing Manager, for Ultra Mobile Group and NAND Solutions
Group, Intel Asia Pacific., explained this performance gap saying, DRAM
is inherently faster than NAND or flash-based SSD in random reads and writes.
Put into an array (they are not a typical SSD), they generally have greater
random I/O performance than even Intels X25-E. However, they are very
expensive (DRAM itself is approx 10x NAND), with system price points reaching
up to $250,000in most applications, a DRAM-based SSD array with battery
backup costs 20-25x more than a comparable NAND SSD. DRAM SSDs have actually
been around for decades and yet the penetration of this technology into enterprise
computing is miniscule due to its price and volatile nature.
- SSDs can be used as server boot devices.
When compared to HDDs, SSDs enable faster boot (typically 30%), consume
lower power, and are more reliable.
- Use SSDs for high throughput, high IOPs,
low latency application storage. If storage I/O is the application bottleneck,
replacing with SSDs shifts the bottleneck back to CPU utilization. Example
applications include video streaming, search query, and OLTP.
- Use SSDs for building a high performance
storage tier. Many applications have hot and cold (or long tail) data.
By creating a storage tier, the solution cost of a deployment can be
reduced significantly.
- Consider SSDs as a lower cost alternative
to placing application data in memory. Many applications create memory-based
databases to achieve low latency access times. For many I/O bound applications,
memory is typically being used as a buffer for disk data. The lower
latency and higher throughput of SSDs promise to require less memory
for buffering while maintaining the quality of service objectives of
the application.
|
High endurance
Flash-based SSDs, on account of the design, wear out after repetitive data writes
due to charges that get trapped in the dielectric oxide layer between the substrate
and the floating gate of the cell. Sen added, The endurance problem in
SSDs can lead to failure of disks as SSDs are prone to wear leveling.
Yiu mentioned that Intel can handle the wear-leveling effectively
so that SSDs do not fail. According to him, Intels wear-leveling technology
is designed as Active Wear-Leveling technology, which is combined with dynamic
wear-leveling and static wear-leveling technology. A traditional wear leveling
scheme has a wear leveling efficiency of about 3X, which means they have about
three times as many writes on the block that has the most writes to it compared
with the blocks that have the average number of writes to it. To compare with
traditional wear-leveling technologies, we could probably safely put about three
times more NAND writes on our SSD while safely remaining within the NAND cycling
capabilities because of the great efficiency that exists in the companys
active wear-leveling algorithms.
Saving power
DRAM-based SSDs usually incorporate internal battery and backup storage systems
to ensure data persistence. Sivashankaran said, DRAM SSDs would increase
the operational costs such as power and cooling. Further, while a spinning disk
would consume 12 Watts of power, a flash SSD would consume only 2.5 Watts of
power. So, flash-based SSDs would help save on operations costs in the long
run.
Sivashankaran said, The ratio of cost of per gigabyte of storage between
HDD and flash SSD is $5:$35 but when SSDs are pitched against DRAM SSDs the
cost equation looks like this$100 for DRAM SSD as against $35 in the case
of flash SSDs. The cost of DRAM SSDs are prohibitively high and hence its applications
are very niche.
Nirmal Puranik, Brand ManagerStorage, Systems and Technology
Group, IBM India/SA agreed, NAND-based SSDs would consume lesser power
when compared to spinning disk. DRAM-based SSDs performance today does not justify
its high costs and hence we see higher and faster adoption of NAND SSDs as we
are seeing a consistent drop in the prices and its getting productized.
Developments in flash SSDs
Yiu explained the advantage of Intel SSDs in enterprise computing is that an
infrastructure already exists using standard SATA and SAS HDDs. Intels
SSDs can drop into these sockets and provide up to a 115x improvement in performance
(which is fast enough for most server acceleration requirements), is more reliable,
and consumes less power than the technology they are replacing (HDD). RAM based
arrays are more complex and require OS and hardware modifications and optimizations,
and again because they are volatile, are only used for accelerating performance
and not for storage.
Suns approach towards increasing the performance of flash SSD is different
and relies on its ZFS file system that combines Solaris Hybrid Storage Pool
technology. Using this technology, which is a part of ZFS, the 7000 family intelligently
and transparently moves data between tiers of read and write optimized SSDs,
DRAM, and disk storage, while managing all of the tiers as a single pool of
storage. Shivashankaran said, A Hybrid Storage Pool approach provides
the benefits of high performance SSDs while still saving money with low cost
high capacity disk drives. Using Hybrid Storage Pool with ZFS, storage architects
can move data from SSDs to SAS or SATA drives seamlessly. He however,
of the opinion that, while SSD performance and operating costs are appealing,
clearly it is not cost-effective in every case to substitute SSDs for mechanical
drives in a storage array. Sun Microsystems is using NAND-based SSD for its
open storage product, Sun Storage 7000 series.
IBM is using two different types of single-level cell (SLC) SSDs, including
drives from STEC and multi-level cell (MLC) technology. In SLC technology, the
limit on the number of writes can be one million, which under some stressed
usage cases could be exceeded in days. For MLC, the number of writes can range
from 10,000 to 100,000 making MLC even 10 to 100 times less durable than SLC.
Puranik said, SLC technology is reliable through wear-leveling, over provisioning
and error correction technologies, but MLC still requires more work and innovation
to bring it into enterprise-class applications.
Narayanaswamy said, I believe NAND SSD offers high scalability and it
can be a fabricated technology in the road ahead for enterprise storage. Give
NAND SSDs another 5-8 years and it would emerge as the default storage technology,
while DRAM SSDs would a remain special purpose, highly scalable technology used
in mission critical application. He added that to get the performance
out from NAND SSDs vendors need to build transactional file system such as Solaris
ZFS that can access multiple discs seamlessly. MindTree too is working towards
developing a new transactional file system that can access and move data from
FCs, to SSDs to SATA/SAS drives.
akhtar.pasha@expressindia.com
|