[erlang-questions] Kai - An Open Source Implementation of Amazon's Dynamo

[Paul Mineiro]

On Mon, 14 Jul 2008, Takeru INOUE wrote:

> Thank you for detailed explanation.
> I understand that tcerl has no disadvantage against dets in architectural point.
> Don't you have any benchmark results?

tcerl appears faster than dets on simple benchmarks (e.g., random inserts
and deletes of random small (e.g, size (erlang:term_to_binary (X)) < 32)
erlang terms), until the scale of the data gets to the point that the O
(log N) loses to the O (1), which occurs around the 10 million record
point, after which dets becomes progressively better at equality lookup. i
say appears because i don't put much stock in such benchmarks, they never
seem to correspond to actual usage.  it was more for me to make sure i
didn't do something horribly inefficient in the driver.

our main motivation for tcerl was ordered_set on disk. we didn't have any
particular performance difficulty with dets for equality queries, and the
file size problem never really bothered us since we heavily use fragmented
tables for availability and load balancing anyway.  we just do alot of
range queries in our particular application.

> To make installation easy, we're planning to support dets or mnesia first.
> But I'd like to support tcerl as next step for better performance.

well, if you target mnesia, you can then use tcerl via the mnesia storage
API extension.

http://code.google.com/p/mnesiaex/

-- p