Hi Warren, all!
Your statement is true ...
Warren Young wrote:
> A lone 2 TB rotating disk will beat a top-of-the-line SSD for linear
> writes, and you can beat an SSD for linear reads with a pair of disks in
> RAID-0 or -1, or four disks in RAID-10. [[...]]
... but irrelevant:
Linear writes and linear reads are not what governs DBMS performance.
More relevant is this:
> [[...]] SSDs have a clearer advantage
> for random I/O, a useful property for databases, but still, you
> shouldn't ignore the fact that SSD writes are expensive.
Especially important is the latency (not throughput!) of random writes
to the log, which may govern your transaction turnaround time.
I don't doubt "SSD writes are expensive", but that holds for any disk
subsystem write (regardless of the technology).
> Therefore, you get the SSD speed benefit only if writes are rare enough
> that more data is coming off the drive at any given time than is being
> written, or if your current disk subsystem is bottlenecked by rotating
> disk head seek time, or some combination.
Exactly: Seek time before writing a commit to the log.
So if your architecture uses a disk subsystem for "stable storage" (as
opposed to MySQL Cluster based on RAM and duplication), its write speed
is a limiting factor for the performance of write transactions.
Joerg Bruehe, MySQL Build Team, joerg.bruehe@stripped
(+49 30) 417 01 487
ORACLE Deutschland B.V. & Co. KG, Komturstrasse 18a, D-12099 Berlin
Geschaeftsfuehrer: Juergen Kunz, Marcel v.d. Molen, Alexander v.d. Ven
Amtsgericht Muenchen: HRA 95603