Product and service reviews are conducted independently by our editorial team, but we sometimes make money when you click on links. Learn more.
 

Advantages Of Flash In The Data Center: Not Just A Flash In The Pan

Advantages Of Flash In The Data Center: Not Just A Flash In The Pan
By

Flash storage is making its way to the data center, offering some unique advantages. But is flash the ultimate storage media, or will some other technology do to flash what flash did to disk?

At Storage Visions 2014, many presentations centered on two themes: huge increases in storage necessitated by higher definition in video and mobile computing, and the predicted trend to put more, or even all, flash memory in the enterprise data center. All presentations seemed to agree on the fact that data center storage would be tiered even more than now, meaning that there would be at least three tiers:

  • Ultra-low or low latency,
  • Low latency/high IOPS,
  • "Cool/cold," long term storage.

Where presenters differed was on which tiers would be all or mostly flash, depending upon which industry segment or company they represented.Fig. 1. San Disk includes 4 tiers, the first tier being ultrafast, expensive Flash as DIMM – a typical use in high frequency trading. Used with permission of SanDisk.

Fig 2.  What type of flash for which workload or tier? Flash memory, and other memory types use much less power, take less space, and generate less heat than spinning disks. They are also considerably faster in terms of latency. 

What is driving increased storage? Video and mobile computing, of course, but in the last few years, those requirements have changed.

As bandwidth and availability increases, memes for example, become more widespread and more imitated. Instead of going to the corner drugstore to play with yo-yos or spin tops, we post high resolution videos on YouTube.

Original Gangnam video on You Tube

6 common resolutions, diff codec versions, diff platforms; total, 2.8 GB for only some of the versions available, one time.

Viewed 1.9 billion times

Copied onto Facebook

~ 52 MB, avg, per instance

~52 MB per occurrence

Flashmob imitations (1st one, 9/12/12) copied and distributed

~66 MB avg., per instance

Popularity widely varies, from millions to a handful of times viewed and/ or copied.

For just one meme, storage requirements are enormous. 

The Flash Trend: Flash In The Data Center

With most of the emphasis on flash, hybrid solutions, and occasional SSD or SSD/hybrid advocates, what is fueling the trend to more flash? 

What is flash compared to other memory? A flash memory is a type of non-volatile memory that is typically erased and reprogrammed in blocks instead of one byte at a time. DRAM memory, the "memory" in most PCs, is volatile, so it requires power to save charges in each cell (memory) compared to flash, which will hold its charge without external power. 

When we speak of storage flash, we mean NAND flash, named for the type of logic gate used. Flash chips, and the transistors that make them up, are similarly sized as DRAM. So why is flash capable of much higher density? Flash chips are essentially 3-D chips - multiple layers on top of each other. Normally, heat would be a problem for DRAM, but with flash, voltages are lower, and because flash charge (memory) can last for years, no electricity is needed for storage, so most layers would not be in use at any given time (aka "dark silicon"). 

But flash is slower than DRAM, thus transfer speeds are slower. Flash bandwidth is also smaller compared to DRAM: with a 100X lower typical speed, and fewer transfers per cycle, flash IOPS lag behind DRAM by about three orders of magnitude, but can have two orders of magnitude greater density.

Why use flash in the data center? 

Flash costs much less than EEPROM and DRAM, and is up to 5x better in terms of dollars per GB. Flash power consumption is up to 40x better than DRAM, or power per GB. Though IOPS (Input/Output Operations Per Second) is a commonly used storage metric, and flash isn't at the top of the list in that regard, flash is 40x better than HDDs in terms of cost per IOPS, and is 600x better than HDDs in terms of power per IOPS.

Cost is determined by cost per wafer, number of dies per wafer, memory density, memory area per die, and patterning density (memory per die per memory density). After 2005 or so, NAND prices dropped below DRAM (used for main computer memory, faster and better addressability than flash, but requires refresh every few ms) prices.

How Flash Is Used In The Data Center

There are three broadly defined use cases for flash in the enterprise data center: all flash arrays, hybrid arrays, and controller based flash.

  • All Flash Arrays (AFA)

All Flash NAND arrays are usually expensive ($/GB), but have the best performance. These arrays frequently feature sub ms latency. 

There are two categories here; the first is a traditional storage array with only SSDs, and tweaked controllers or software to support flash wear leveling and latency. Second is the all flash array, a storage array engineered to support flash and low latency, are scalable to very high performance, data protection (RAID), and efficiency via deduplication and compression. 

  • Hybrid (HDD/Flash) Arrays

Hybrid arrays are a good compromise: 2 to 5 percent of storage capacity may double the IOPS, and reduce latency. In comparing a 64 TB fibre channel array (224 FC drives) vs a hybrid array (56 FC drives, 16 TB flash), the cost is 50 percent lower and the hybrid approach yields 67 percent power savings and 67 percent space savings.The latency is reduced from about 4.5 ms to 3.5ms, for the same number of IOPS.  

  • Controller Based Flash

Flash may also be added to controllers, either in storage or on the network. This automatically places data where access is fastest, and provides more throughput in a disk storage system without extra high performance hard drives. Flash may be added to boost particular hardware or software, in a scenario named host-based flash, where the flash is on a PCI card on the host system. This acts as tier 0 storage in front of networked storage.

The three cases where this can be used include: shared storage OS in a hosted VM, in high-performance direct attached storage, and a shared RAID storage subsystem in a hosted VM. Fig. 3. Common Flash Usage Scenarios. Provided courtesy of SNIA, and the author, Marty Czekalski, President, SCSI Trade Association, and Sr. Staff Program Manager, Seagate Technology. 

What's Next?

Storage class memory is coming, defined as any or a mix of new flash types, within the next few years, and will significantly alter how we view storage. In the meantime, faster, cheaper, less power hungry, and smaller storage arrays will populate data centers. That they'll do so with tricks to maximize lifetimes with inexpensive NAND flash is beside the point. 

The advantages these features yield for day-to-day data center operations will allow data center managers to keep up with increasing storage requirements without building out data centers, yielding greater access speeds with lower power requirements.