Robin points out some very valid observations including the unfortunate view that SSDs are just like rotating storage only faster and by just adding more disks can solve their performance problems at a lower price compared to implementing SSDs. Often this is not the case. Here’s just one example of why.

To get lots of IOPs out of a disk farm, the server and software need to be able to develop a large queue. You can visualize this if you assume for a moment a single threaded process asks for random data from a single disk. If the disk has to move its read/write head, the disk’s not going to complete this IO for 5 to 10 milliseconds. This is eons in processor time and would limit the system to about 100 to 200 IOs per second. So in order to get thousands of IOPS (aggregated) you clearly need many spindles. But more importantly, the server and the app also need to be able to generate the additional, simultaneous parallel processes or threads (the queue), each of which will remain open for the 5 to 10 milliseconds. Hence the server’s queue capacity and the software’s ability to generate parallel requests are integral in getting high IO rates when using disk arrays. In normal operation there may be hundreds or thousands of open processes waiting for service by the disk farm.

With SSDs this is not the case. Even a single thread or process will get serviced in microseconds, so potentially, 10’s of ‘000s of IOPS can be achieved with little or no queue in the server.

There are many benefits to this.

Fewer, Cheaper Servers with lower system power consumption — some of which is due to lower server consumption, some due to the lower consumption of the SSD. You also get lower licensing and administration costs, less tuning and less issue with fragmentation, but often the biggest cost saving is in server stability and preventing server performance spikes or worse case, crashes. This is not an intuitive benefit of SSDs unless you examine the queue issue.

With a disk farm under a high transaction load we’ve established that you have a big queue. This uses lots of server resources and because of the long delays of the drives, any large burst of requests will initially get staged in the server which can increase the queue beyond the server capacity causing instabilities or a crash.

With SSDs, IOPs are handled synchronously and at about the same speed as the network or server can supply them, so crashes due to requests backing up in the server are eliminated.

This example points out the need to analyze the costs of SSDs in conjunction with the rest of the system. Only then can the significant benefits be measured against the higher costs per byte.

SSD has been kicking around for a long time and has been a real ROI winner for applications where low storage IO latency is critical. Right now, we are seeing an explosion in solid state memory technology - manufacturers are designing for vertical applications, including Personal media devices, set-top boxes, the mobile phones, cameras, data center and laptops. Exciting stuff !!!

Today’s bang for the buck for the PC industry is laptops with SSD’s. Fast boot times, power-on data availability, no spin-up required. In comparing SSD to spinning media, disk performance for laptops w/ 5400 rpm drives get about 25 MB/s on a good day. The new flash memory chips read data at about 60-80MB/s with latency of 5uS vs. 180 uS for laptop disks. So, SSD does have the performance capability for laptops and without the disk spin-up time. I can live with that…

Hybrid drives are here to stay for a while, I put a hybrid drive in my HP laptop, its great… As soon as the flash or the next solid state memory technology (Spansion’s products) improves write capabilities and reduces price, its good bye to spinning disk drives on laptops.

There is a rumor that SSD with increased write capabilities will reach price parity with spinning media in 5-7 years. I believe broad adoption of SSD will occur for economic reasons. Consumer products will move to SSD when the costs of SDD out ways the cost of warranty repairs and support calls for spinning disk failures or quality of service related issues (the device stalls due to seek problems).

The enterprise will move to SSD when they examine the costs in lost revenue due to degraded raid performance and when they compare data center costs cause by failed drives. The enterprise has additional facility energy costs related to spinning drives and cooling the data center.

I hope my math is right here..
32 spinning 15K 3.5″ drives (500W) costs about $2500 to run and cool in well designed data center - more $$ in a poorly designed data center. In Japan, it costs about $5000 for the same drives.

Depending on drive capacity (300GB vs 72GB), spinning enterprise drives burn between 43mW and 210mW per GB, costing between $7.10 and $35 per GB annually in energy cost. Drive specs show SATA drives use about 5-15% less power than SAS, fibre channel and scsi. I’m not sure I believe that number. Laptop drives are use about 20mW/GB. Sandisk claims 15mW/GB for their 64GB SATA Flash disk, costing about $2.10 annually in data center energy per GB. Too bad flash has such limited write capabilities.

On the optical front, Optware had already produced a 30GB holographic memory card for $0.90 and InPhase has a 300GB 5.25″ disk, not R/W, but worm drives. The problem with this technology today the drive is over $1000. LOTS, the pioneer in optical “write once” tape closed its doors a few years ago.

I think we have a ways to go before R/W optical media become a reality, not in my life time anyway