[erlang-questions] Re: Best way to implement a simple cache

[Matt Stancliff]

On Nov 12, 2009, at 5:17 PM, Thijs wrote:

> On Nov 13, 8:36 am, Scott Lystig Fritchie <fritc...@REDACTED>
> wrote:
>> ... except that public/protected ETS table access isn't free, either.
>> The VM has to play games with controlling multi-scheduler access to
>> ETS's underlying hash table/binary tree.  I haven't measured with an
>
> You are right, but in my experience a public ETS table significantly
> outperforms process-mediated access. I'm curious how did you optimize
> this further?

   Let's solve this with numbers!

Test                    ||  Multiplier  || Percent Slowdown || Reads  
Per Second
= 
= 
= 
= 
= 
= 
= 
========================================================================
ets plain                       1.00x           0%                   
553,384.3
ets public                      1.01x           1%                   
550,329.3
ets public named table          1.10x          10%                   
502,504.7
mnesia ram copies               1.87x          87%                   
296,011.2
process based cache             9.86x         886%                    
56,141.6
BDB driver                     56.37x       5,537%                     
9,816.3

While I have my benchmarking hat on, let's check function calls too:
Call Method             ||  Multiplier  || Percent Slowdown || Calls  
Per Second
= 
= 
= 
= 
= 
= 
= 
========================================================================
apply(?MODULE, foo, [])        1.00x            0%                  
96,089,747.8
?MODULE:foo()                  1.01x            1%                  
94,809,477.4
foo()                          1.09x            9%                  
88,485,052.9
fun() -> ok end()              3.15x          215%                  
30,489,431.6
apply(foo, [])                 3.50x          250%                  
27,481,044.5
Module:Fun()                   5.22x          422%                  
18,392,358.4

The slowest method of function invocation still yields 18+ million  
calls per second.


Notes:
   • Tests were run on my ancient (almost four year old) iMac (Core  
Duo 2GHz):
Erlang R13B02 (erts-5.7.3) [source] [smp:2:2] [rq:2] [async-threads:0]  
[kernel-poll:false]
   • No scheduler binding to CPUs.
   • For the function call benchmark: foo() -> ok.  Module = ?MODULE,  
Fun = foo.
   • The "process based cache" is my fleshed out production version of  
the duomark process cache.

Disclaimer: The results vary in absolute numbers for multiple runs of  
the benchmark,
but results retain their relative positions in the rankings each time.
I'll get around to making a statistically valid erlang benchmark suite  
one day.


-Matt
-- 
Matt Stancliff                    San Jose, CA
AIM: seijimr              iPhone: 678-591-9337
"The best way to predict the future is to invent it." --Alan Kay