loggerhead vs packs performance

[Michael Hudson]

Robert asked me to do check loggerhead's performance when pointed at  
a repository using packs (I tested with was bzr.dev).

The performance is summarized in this table (the numbers are the  
minimum page load time in ms reported by 'ab -n 15' on the given view):

+---------------+---------------+---------------+--------------- 
+---------------+
|               |  changelog    |  inventory    |  annotate     |   
revision     |
+---------------+---------------+---------------+--------------- 
+---------------+
| pack_cached   |    1371       |    3174       |    23259      |     
10606      |
+---------------+---------------+---------------+--------------- 
+---------------+
| knit_cached   |    1373       |    2174       |    16625      |      
4135      |
+---------------+---------------+---------------+--------------- 
+---------------+
| pack_nocache  |    6760       |    4169       |    24439      |     
10178      |
+---------------+---------------+---------------+--------------- 
+---------------+
| knit_nocache  |    5645       |    3530       |    19215      |      
4249      |
+---------------+---------------+---------------+--------------- 
+---------------+

(this was done on my old slow os x powerbook, which is why the  
numbers are SO uniformly bad).

For the changelog and inventory views, packs are a little slower, but  
probably nothing some tuning and effort wouldn't sort out.  Annotate  
is pretty slow (but that may have been expected?  And the rendering  
performance is still dreadful here).  The revision view, though, is  
atrociously slow with packs, and this is borne out with timings of  
'bzr diff':

[mwh at quickbeam tmp]$ (cd $BZR_DEV_PACKS; time bzr2 diff -r 2859..2860  
 > /dev/null)

real    0m12.344s
user    0m8.996s
sys     0m0.950s
[mwh at quickbeam tmp]$ (cd $BZR_DEV_KNITS; time bzr2 diff -r 2859..2860  
 > /dev/null)

real    0m4.151s
user    0m2.651s
sys     0m0.685s

Cheers,
mwh