Hacker News new | past | comments | ask | show | jobs | submit login
Where SSDs Don't Make Sense in Server Applications (mvdirona.com)
21 points by neilc on April 12, 2009 | hide | past | favorite | 10 comments



A 2008 Sun article sees Flash as having the following position in the server storage hierarchy: cache, DRAM, Flash, disk, tape: http://blogs.sun.com/jonathan/entry/not_a_flash_in_the

Which is what the present article is talking about in this line:

The benefits of SSDs as an intermediate caching tier are also limited, and the cost of provisioning such a tier was justified for fewer than 10% of the examined workloads


I guess the real questions is are SSDs good enough to replace data that now sits in RAM? For example, Google keeps their entire search index in RAM. Would they be better off keeping it in SSDs?


Not even close. The bandwidths of RAM vs. SSDs differ by a couple orders of magnitude.


Especially with the Intel X-25 at around $4.50 per GB and DDR2 running at $10 per GB - there just isn't much of a cost savings for completely disparate performance.


When you calculate the price of RAM, consider the slots as well. Putting a TB of RAM in a server requires more than just DIMMs.


But as the paper linked by the article[1] points out, a single server can hold much more data in SSD storage then in onboard RAM.

For ordinary databases, one sweet spot for SSDs is for index tablespaces.

[1] - http://research.microsoft.com/en-us/um/people/antr/ms/ssd.pd...


Shouldn't he wait until the cost comes down and settles a bit before he bothers writing a whole screed about the expense? It's like decrying the HDTV as not worth the money -- they're over $10,000 each! Or, they were, a few years ago.


Well, obviously the price of SSDs will change in the future, but the question is whether they are a good idea for server workloads now, not in 2 or 3 years.


The paper says "For each workload, we also calculated the break-even capacity/dollar at which the best [SSD cache] solution would equal the cost of a disk-based solution." Their results tell you exactly how cheap SSDs need to get before they're worth using. Given that SSDs have improved 5x-10x on several metrics since this paper, they are already suitable for some workloads.

Also, the paper finds that for many of their workloads performance effectively does not matter so SSDs will only be cheaper if they are cheaper per GB than disks, which will probably never happen. They're not saying that SSDs are too expensive; they're saying that SSDs will always be too expensive. It's not clear to me that this conclusion is correct, but it is interesting.


Writing about what performance may or may not look like in 3 years time after changes that may or may not happen seems like a waste of time.




Join us for AI Startup School this June 16-17 in San Francisco!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: