01:39 PM
George Crump
George Crump

Quick Guide To Flash Storage Latency Wars

Because latency is the key performance differentiator in server-side flash, SSD, PCIe and memory bus flash storage vendors are competing on speed.

Server-side flash is all about performance, and the key performance differentiator is latency. Latency is the time required to complete a transaction between a host and a storage system, such as a read or write operation. As a result, a "latency war" has broken out in the flash community about the best way to reduce flash latency.

Latency actually has several levels. The first level is the time it takes the media to position and be ready to respond to an I/O request. With a hard drive, this latency was the end of the discussion, as the milliseconds it took for a hard drive to get in position overshadowed any other latency in the storage communication chain.

Solid state disk (SSD) drives -- which are flash inside a hard drive container -- eliminated that device latency. SSDs can put themselves in position instantly since they have no moving parts, no platters to rotate. But that exposed other areas of latency in the storage protocol stack. For example, the time it takes for the I/O to work its way through the overhead of SCSI became noticeable.

This led to the introduction and rapid adoption of PCIe-based flash. Most of these cards eliminated the storage protocol stack altogether. The communication was direct to the application or operating system over the PCIe bus. Eliminating the storage protocol stack meant that special drivers had to be created for the various operating systems that a data center might have.

Vendors then introduced API sets that would allow users to write directly to the PCIe flash card from within the application for a further reduction in latency, not only avoiding the storage I/O stack but also avoiding the operating system itself.

What PCIe began to show, though, was that there was another level of latency. The PCIe bus itself. The PCIe bus routes through a motherboard-based fabric that allows multiple cards to share PCIe bandwidth. Some higher end servers will have multiple PCIe hubs in order to better route data. But even a high-end server when burdened with lots of I/O can become PCIe bottlenecked and lead to latency. (Remember that PCIe supports more than storage I/O.)

The next tier in latency elimination is to eliminate the PCIe bus altogether. Several vendors are introducing memory-bus-based flash storage. These flash storage devices come in a memory DIMM form factor and can act as storage with a device driver, similar to a PCIe SSD. Even more interesting, with a tweak to server system BIOS, it can act as main memory to the server. Imagine 400 GBs of "RAM" via flash on a single DIMM. Using the memory bus provides even more I/O channels and greater bandwidth; it was, after all, designed to support DRAM.

In both implementation modes, the use cases are very interesting. The ability to create very dense servers with terabytes (if not petabytes) of flash capacity in a 1U system changes the data center design game quite a bit.

There is no single perfect solution, as SSD, PCIe flash and memory flash all have their ideal use cases, and many data centers may have a mixture of all three. Designing applications for extreme high performance with almost zero latency is now reality. But their use is not limited to the performance fringe. For example imagine using these technologies to design a single physical server to support 10,000 plus desktops. We have the processing power available to us, but the latency of storage is no longer the roadblock it always was.

Comment  | 
Print  | 
More Insights
Threaded  |  Newest First  |  Oldest First
User Rank: Apprentice
10/31/2013 | 2:25:02 AM
re: Quick Guide To Flash Storage Latency Wars
Great summary describing latency in storage. It would be good to have a table comparing latencies of HDDs, SSDs, PCIe and flash memory DIMMs.
Joe Stanganelli
Joe Stanganelli,
User Rank: Ninja
10/31/2013 | 3:58:11 AM
re: Quick Guide To Flash Storage Latency Wars
Of course, this all stands to change over the next several years as SSDs gradually become obsolete (except, possibly, in hybrid setups) under Moore's Law. http://www.enterpriseefficienc...
D. Henschen
D. Henschen,
User Rank: Apprentice
10/31/2013 | 10:49:46 AM
re: Quick Guide To Flash Storage Latency Wars
I heard an enlightening cost comparison the other day by Ari Zilka, CTO at Hortonworks. 1Terabyte of RAM = $70,000; 1TB of Flash = $8,000 to $20,000, depending on quality/discounts; 1 TB of hard drive = $60 to $100 depending on quality/performance. Maybe somebody would quibble with the exact figures, but it's clear there is a cost to real-time performance.
User Rank: Apprentice
10/31/2013 | 11:25:23 AM
re: Quick Guide To Flash Storage Latency Wars
Any review on Tape storage,as i don't think it going away any time soon, as it is still the most cost-effective and simplest way to archive huge amounts of data.
Register for Dark Reading Newsletters
White Papers
Current Issue
E-Commerce Security: What Every Enterprise Needs to Know
The mainstream use of EMV smartcards in the US has experts predicting an increase in online fraud. Organizations will need to look at new tools and processes for building better breach detection and response capabilities.
Flash Poll
10 Recommendations for Outsourcing Security
10 Recommendations for Outsourcing Security
Enterprises today have a wide range of third-party options to help improve their defenses, including MSSPs, auditing and penetration testing, and DDoS protection. But are there situations in which a service provider might actually increase risk?
Twitter Feed
Dark Reading - Bug Report
Bug Report
Enterprise Vulnerabilities
From DHS/US-CERT's National Vulnerability Database
Published: 2015-10-15
The Direct Rendering Manager (DRM) subsystem in the Linux kernel through 4.x mishandles requests for Graphics Execution Manager (GEM) objects, which allows context-dependent attackers to cause a denial of service (memory consumption) via an application that processes graphics data, as demonstrated b...

Published: 2015-10-15
netstat in IBM AIX 5.3, 6.1, and 7.1 and VIOS 2.2.x, when a fibre channel adapter is used, allows local users to gain privileges via unspecified vectors.

Published: 2015-10-15
Cross-site request forgery (CSRF) vulnerability in eXtplorer before 2.1.8 allows remote attackers to hijack the authentication of arbitrary users for requests that execute PHP code.

Published: 2015-10-15
Directory traversal vulnerability in QNAP QTS before 4.1.4 build 0910 and 4.2.x before 4.2.0 RC2 build 0910, when AFP is enabled, allows remote attackers to read or write to arbitrary files by leveraging access to an OS X (1) user or (2) guest account.

Published: 2015-10-15
Cisco Application Policy Infrastructure Controller (APIC) 1.1j allows local users to gain privileges via vectors involving addition of an SSH key, aka Bug ID CSCuw46076.

Dark Reading Radio