hot path optimizations in uma_zalloc() & uma_zfree()

[John Baldwin]

On Wednesday 29 June 2005 06:08 pm, ant wrote:
> Hi folks.
>
> I just tryed to make buckets management in perCPU cache like in
> Solaris (see paper of Jeff Bonwick - Magazines and Vmem)
> and got perfomance gain around 10% in my test program.
> Then i made another minor code optimization and got another 10%.
> The program just creates and destroys sockets in loop.
>
> I suppose the reason of first gain lies in increasing of cpu cache hits.
> In current fbsd code allocations and freeings deal with
> separate buckets. Buckets are changed when one of them
> became full or empty first. In Solaris this work is pure LIFO:
> i.e. alloc() and free() work with one bucket - the current bucket
> (it is called magazine there), that's why cache hit rate is bigger.
>
> Another optimization is very trivial, for example:
> -   bucket->ub_cnt--;
> -   item = bucket->ub_bucket[bucket->ub_cnt];
> +   item = bucket->ub_bucket[--bucket->ub_cnt];
> (see the patch)

This produces no change in the object code though as gcc is smart enough to 
implement both cases the same.

I ran ministat against your tests with 1000 sockets loop and there isn't a lot 
of difference in the user times:

x one.u
+ two.u
* three.u
+--------------------------------------------------------------------------+
|*              *      *       x      x       +              x      ++     |
| |__________A__M________|  |_________M____A____|__________|_A______M_____||
+--------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   3         0.085         0.117         0.093   0.098333333   0.016653328
+   3         0.101         0.125         0.124    0.11666667   0.013576941
No difference proven at 95.0% confidence
*   3         0.054         0.077          0.07         0.067   0.011789826
No difference proven at 95.0% confidence

There is some difference in the system times though:

> /usr/src/tools/tools/ministat/ministat one.s two.s three.s
x one.s
+ two.s
* three.s
+--------------------------------------------------------------------------+
|*  *  *    ++  +                                         x       x       x|
||__A__|    |MA_|                                         |_______A_______||
+--------------------------------------------------------------------------+
    N           Min           Max        Median           Avg        Stddev
x   3         2.689         2.797         2.739     2.7416667    0.05404936
+   3         2.373         2.403         2.381     2.3856667   0.015534907
Difference at 95.0% confidence
        -0.356 +/- 0.0901334
        -12.9848% +/- 3.28754%
        (Student's t, pooled s = 0.039766)
*   3         2.298          2.34          2.32     2.3193333   0.021007935
Difference at 95.0% confidence
        -0.422333 +/- 0.0929396
        -15.4043% +/- 3.38989%
        (Student's t, pooled s = 0.0410041)

There is also some difference between the second and third runs though not 
huge.  Given that your ++ and -- optimizations don't produce any actual 
changes in the compiled code I think that is more due to your testing 
environment.  A much better test might be to reboot into single user mode and 
run the socket loop program with 10000 or 100000 sockets and run that in a 
loop 20 times for each case and log the output.  That should give you some 
better benchmarks with less noise from your environment.  If you haven't used 
ministat before (/usr/src/tools/tools/ministat), it's a neat little tool.  
You just hand it data files that contain one number per line and it does the 
work for you.  Also, given that here are SMP implications here, it might be 
fruitful to have another benchmark which spawned several threads (one per CPU 
maybe?) and had each of them sit in the same loop.

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org