Implementing Garbage Collection forActive Objects on Top of ErlangSigmund HansenMaster thesis autumn 2014

Problem

I ABS is a modeling language for distributed systemsI ABS has an Erlang back endI Back end cannot collect garbage processes

GoalsThe garbage collector should be:

I FastIt should not considerably slow down simulations

I ComprehensiveIt should collect most or all garbage

I CorrectIt should not collect resources required later

ABS Overview

I Executable modelsI Functional subsetI Object-orientedI Concurrency:

I Active objectsI Scheduling in Concurrent Object Groups (COGs)I RPC with message passing

Erlang Overview

I Concurrency:I Actor modelI Message passing

I Fault-tolerantI Functional

Garbage Collection OverviewAutomatic memory management for dynamically allocatedmemory.

I Reference countingI Tracing

I Object graph traversalI De-allocation of unmarked objects

Active Object Collection

I Kept alive by associated processI Must kill the processI High-level implementation

Choice of Algorithm

I Reference counting cannot collect cyclic garbage.I Killing processes, not collecting memory,

i.e. no moving collectorsI Greater risk with non-standard collection algorithms

Mark and sweep that stops the world implemented.

Stopping the World

I No interruptsI Stop by messagingI Tasks and scheduling must stop

Blocking Operations

I Blocking on futuresI Blocking on object instantiationI Inappropriate COG state during block

COG State Machine (Before)

Not running

Running None found

[runnable task found]token,task error

[no runnable task found]new task,task state changed,task error

new task,task state changed

COG State Machine (After)

Not0running

Stopped

Resume0state

Running

None0found

Blocked

Blocked0task

ok

stop0world

[runnable0task0found]token,task0error

[no0runnable0task0found]

new0task,task0state0changed,task0error,inc/dec0reference0count

stop0world

resume0world[reference0count0!=00,Resume0state0=0Not0running]

stop0world

new0task,task0state0change,inc/dec0reference0count

stop0world

running0task0statechanged0to0blocked

blocked0task0statechanged0to0runnable

stop0worldresume0world[reference0count0!=00,Resume0state0=0Blocked]

new0task,task0state0change,inc/dec0reference0count

get0references,inc/dec0reference0count

resume0world[reference0count0=00]

Stopping Running Tasks

I Will reduce waitingI COG messages taskI Task then blocks

Asynchronous Method Calls onInactive Objects

I Task initialization is synchronousI Waits until object is activatedI Respond to GC messages while waiting

Tasks CreatedWhile the World is Stopping

I Task is created by a futureI Stopped COGs do not receive task creation messagesI Futures temporarily become roots and keep parameters

Future State Machine (Before)

Loop

Result

init

Await_start

Parameters

Await_completion

Started

Taskcompleted

Taskerror

Future State Machine (After)

Wait

Loop

Result

init

Await_start

Parametersget_references

Started

Startedget_references

Taskcomplete

Taskerror

die

A Complete Viewof COGs and FuturesWhy is it required?

Marking Phase

I Messages all grays in parallelI BN+1 = BN ∪ GN

I GN+1 = References(GN) \ BN+1

I Sweeps when gray set is empty

Sweeping

I Identifies white objectsI Identifies white futuresI Sends message to kill themI Resume message sent to COGs

Collecting COGs with ReferenceCounting

I Counts objects left in groupI COG stops if world resumes and it is emptyI Messages garbage collector

Static analysis

I Types that may contain referencesI Reaching unawaited futuresI Other potential analyses:

I Loops with synchronization pointsI Deep functions

Triggering Garbage CollectionWhen should garbage be collected:

I On time?I Number of objects and futures?I Process ratio?

Data Collection

I Garbage collector state changesI Number of objects and futures sweptI Memory usage and number of objects, futures and COGs

Test Cases

I Ping pongBasic test with cyclic garbage

I SequencesUnstoppable loop

I Prime SieveLong-running tasks

I MapReduce - IndexingReal-world test case

Ping Pong - Time

Collection scheme Real time Relative Relative without idle

Before GC 1274 ms 100.0 % 100.0 %Never collect 1257 ms 98.6 % 93.5 %Always collect 1854 ms 145.5 % 311.3 %Collect on count 1293 ms 101.4 % 106.7 %Collect on time 1261 ms 98.9 % 94.9 %+ Stop running 1265 ms 99.2 % 96.5 %Collect on either 1290 ms 101.2 % 105.5 %+ Stop running 1298 ms 101.8 % 108.4 %

Ping Pong - No GC

0

250

500

750

14

16

18

TotalM

emory in M

iB

0 ms 100 ms 200 ms 300 ms 400 ms 500 msTime

variable

cogs

objects

futures

Ping Pong - Always collecting

0

100

200

300

12.5

15.0

17.5

20.0

22.5

TotalM

emory in M

iB

0 ms 1220 ms 2440 ms 3660 ms 4880 ms 6100 msTime

variable

cogs

objects

futures

Ping Pong - By count

0

20

40

60

13.5

14.0

14.5

15.0

15.5

TotalM

emory in M

iB

0 ms 140 ms 280 ms 420 ms 560 ms 700 msTime

variable

cogs

objects

futures

Ping Pong - Timed

0

50

100

150

200

13.5

14.0

14.5

15.0

15.5

16.0

16.5

TotalM

emory in M

iB

0 ms 100 ms 200 ms 300 ms 400 ms 500 msTime

variable

cogs

objects

futures

Infinite Ping Pong - No GC

0

50000

100000

150000

0

300

600

900

1200

TotalM

emory in M

iB

0 ms 31120 ms 62240 ms 93360 ms 124480 ms 155600 msTime

variable

cogs

objects

futures

Infinite Ping Pong - Timed

0

2000

4000

6000

8000

50

100

TotalM

emory in M

iB

0 ms 11900 ms 23800 ms 35700 ms 47600 ms 59500 msTime

variable

cogs

objects

futures

Sequences - Time

Collection scheme Real time Relative Relative without idle

Before GC 1395 ms 100.0 % 100.0 %Never collect 1439 ms 103.2 % 111.2 %Always collect 39159 ms 2807.2 % 9662.1 %Collect on count 16801 ms 1204.4 % 4001.0 %Collect on time 1622 ms 116.3 % 157.5 %+ Stop running 1530 ms 109.7 % 134.1 %Collect on either 17054 ms 1222.6 % 4065.1 %+ Stop running 12614 ms 904.3 % 2940.8 %

Sequences - No GC

0

1000

2000

3000

4000

5000

20

30

40

50

60

70

TotalM

emory in M

iB

0 ms 720 ms 1440 ms 2160 ms 2880 ms 3600 msTime

variable

cogs

objects

futures

Sequences - Timed

0

1000

2000

3000

4000

5000

20

30

40

50

60

TotalM

emory in M

iB

0 ms 860 ms 1720 ms 2580 ms 3440 ms 4300 msTime

variable

cogs

objects

futures

Sequences -Timed stopping running processes

0

200

400

600

16

20

24Total

Mem

ory in MiB

0 ms 700 ms 1400 ms 2100 ms 2800 ms 3500 msTime

variable

cogs

objects

futures

Prime Sieve - Time

Collection scheme Real time Relative Relative without idle

Before GC 1448 ms 100.0 % 100.0 %Never collect 1453 ms 100.3 % 101.1 %Always collect 1903 ms 131.4 % 201.6 %Collect on count 1482 ms 102.4 % 107.6 %Collect on time 1454 ms 100.4 % 101.3 %+ Stop running 1542 ms 106.5 % 120.9 %Collect on either 1472 ms 101.6 % 105.3 %+ Stop running 1545 ms 106.7 % 121.6 %

Indexing - Time

Collection scheme Real time Relative Relative without idle

Before GC 1356 ms 100.0 % 100.0 %Never collect 1373 ms 101.2 % 104.7 %Always collect 5043 ms 371.8 % 1134.2 %Collect on count 1408 ms 103.9 % 114.7 %Collect on time 1398 ms 103.1 % 111.7 %+ Stop running 1405 ms 103.6 % 113.9 %Collect on either 1410 ms 104.0 % 115.3 %+ Stop running 1434 ms 105.7 % 121.9 %

ConclusionTimed trigger gives acceptable results.

I Acceptably fastI Mark and sweep is completeI Measures taken stopping the world to ensure correctness

DiscussionTesting with infinite loops have discouraging results. Futurework to lower overhead may be needed:

I Optimize stopping the worldI Optimize markingI Concurrent collectionI Change of algorithm