544695 - Garbage collector policy should not be the same at all heap sizes

Status: VERIFIED FIXED

(Tamarin Graveyard :: Garbage Collection (mmGC), defect, P2)

Description

[Lars T Hansen]

The garbage collector policy mainly affects non-reference-counted memory and as such it can be seen as analogous to the policy for the "old heap" in a generational system.  There isn't normally a lot of performance pressure on that heap, but there is memory control pressure: While a 2x load factor is fine in a small heap it is too much in a large heap.  For truly large heaps a factor as low as 1.1 may be appropriate.  We should have a policy that provides breathing room for smaller programs but that run the incremental gc fairly aggressively in large heaps, so that large heaps don't grow excessively.  

(Adobe internal context: WE #2534231 and WE #2519943.)

Comment 1

[Lars T Hansen]

Attachment: Patch

Changes GCHeapConfig::gcLoad from a single value to a short table, with hopefully suitable heap size cutoffs for each, introduced as GCHeapConfig::gcLoadCutoff.

Changes GCPolicyManager::adjustL to search the short table.  (Only called at the end of a major GC cycle, so not a performance issue.)

Changes avmshell.cpp to allow the table of load factors and cutoffs to be provided, and moves the definition of the "X" parameter (max multiplier) to a new argument.

Comment 2

[Tommy Reilly]

Comment on attachment 425677 [details] [diff] [review]
Patch

Cool, this makes a lot of sense

Comment 3

[Lars T Hansen]

tr-argo changeset:   3675:c85fe7340d1e

Comment 4

[Felix S. Klock II [:pnkfelix, :fklock]]

Food for thought:

Here are the default load factors, fed explicitly on the command line:
% ./shell/avmshell.r5842 -load 2.5,10,2.0,25,1.75,50,1.5,75,1.25,100,1.1,150,1.05 ../../distilled3.abc 
tf1: 28060 ms - callLater test minmem: 2052 K maxmem: 615552 K

If I peel off that 1.05, so that 1.1 becomes the max inverse load factor:
% ./shell/avmshell.r5842 -load 2.5,10,2.0,25,1.75,50,1.5,75,1.25,100,1.1 ../../distilled3.abc 
tf1: 19923 ms - callLater test minmem: 2052 K maxmem: 628060 K

And peeling off another layer:
% ./shell/avmshell.r5842 -load 2.5,10,2.0,25,1.75,50,1.5,75,1.25 ../../distilled3.abc 
tf1: 13504 ms - callLater test minmem: 2052 K maxmem: 648668 K

(The point being that we don't need to go all the way back to an inverse load factor of 2 to bring performance back for this benchmark.)

Running the shell is fun and all, but my plan for now is to scribble out some formulas and see if I can come up with a disciplined argument for or against them.

Comment 5

[Felix S. Klock II [:pnkfelix, :fklock]]

(In reply to comment #4)
> Food for thought:

oops, that was supposed to go on Bug 619885.