5. Tuning the Memory Pool System for performance

Note

When developing a benchmark to profile your program against, bear in mind that the benchmark should allocate several times the amount of physical memory that you expect to be available to the process. If the total allocation fits into the available memory, there’s no point running a garbage collector at all: you might as well just allocate and never collect.

The most important aspect of tuning the MPS is to choose good sizes for the generations in your generation chain. The ideal size of a generation should be such that when it is collected, most of the blocks allocated in that generation should be found to be dead (and so the cost of scanning and copying them can be avoided entirely). If a generation is collected when its blocks are mostly alive, that is a waste of time.

In the tables below I give the execution time of test-leaf.scm in the toy Scheme interpreter under different settings for its generation chain. (This test case allocates hundreds of millions of small short-lived objects.)

First, the effect of varying the capacity of a chain with a single generation.

Capacity	Mortality	Execution time (user+sys)
100	0.80	362.6
200	0.80	354.9
400	0.80	349.7
800	0.80	314.4
1600	0.80	215.7
3200	0.80	94.0
6400	0.80	53.5
12800	0.80	79.6
25600	0.80	77.6

Second, the effect of varying the mortality of a chain with a single generation.

Capacity	Mortality	Execution time (user+sys)
6400	0.20	55.4
6400	0.40	54.0
6400	0.60	54.0
6400	0.80	53.5
6400	0.99	54.8

Third, the effect of varying the number of generations (all generations being identical).

Generations	Capacity	Mortality	Execution time (user+sys)
1	6400	0.80	53.5
2	6400	0.80	42.4
3	6400	0.80	42.1
4	6400	0.80	42.2
5	6400	0.80	42.2

These tables suggest that:

1. The improvement in performance to be gained by getting generation sizes right is dramatic: much bigger than the small improvements to gained from other techniques.
2. The predicted mortality doesn’t make much difference to the overall execution time (it does affect the distribution of pause times, however: see Scheduling of collections.)
3. You can make generations too big as well as too small.
4. There are rapidly diminishing returns to be gained from adding generations.

Note

Telemetry can be used to discover when generations are being collected and what proportion of blocks were found to be alive.

The table below shows the effect of varying the initial allocation of address space to the arena (using three generations each with capacity 6400 kB, mortality 0.80).

Address space	Extensions	Collections	Execution time (user+sys)
2	32	371	52.0
4	21	370	47.0
8	0	[1]	[1]
14	0	[1]	[1]
16	0	2436	160.5
18	0	1135	89.1
20	0	673	60.6
22	0	484	48.7
24	0	400	43.1
32	0	368	41.2
64	0	368	43.1
128	0	368	46.4
256	0	368	46.3
512	0	368	49.3
1024	0	368	42.0
2048	0	368	43.2
4096	0	368	43.5
8192	0	368	46.1
16384	0	368	49.2
32768	0	368	57.1
65536	0	368	71.1
131072	0	368	101.3
262144	0	368	161.3
524288	0	368	273.0
1048576	0	368	504.6

Note

[1]	(1, 2, 3, 4) With this initial allocation of address space, the test case failed to run to completion after thousands of seconds and tens of thousands of garbage collection cycles.

The lesson here is that the allocation of address space has to be comfortably larger than the working set of the program, but that a very large address space is ruinous to performance.