Mark DURING Sweep rather than Mark then Sweep

1
Mark DURING Sweep rather than Mark then Sweep

• Presented by Ram Mantsour
• Authors
• Chrisitan Queinnec, Barbara Beaudoing,
  Jean-Pierre Queille.

2
Introduction

• Mark DURING Sweep is intended as a real-time GC.
• Regular GCs introduces delays which may affect
  real time applications.
• Features parallel collection algorithm
  (concurrent marker sweeper and application)
• Designed for real-time applications in embedded
  systems.

3
What Mark During Sweep has to fulfill

• Short and predictable response time
• very short time lag between events and handler
  (approx 50 micro sec).
• Guaranteed throughput
• To avoid unexpected GCs.
• Operation with Narrow memory.
• Execution time upper bound for an allocating
  task.

4
Problems with Known schemes

• Stop-and-Copy wastes memory and has problem with
  VM systems.
• Mark-and-sweep in stop-and-collect mode even if
  introduces lazy sweep - has marking delay.
• On-the-fly (concurrent mutator and collector) -
  possible mutator starvation.

5
Mark-During-Sweep - terminology.

• Roots in their usual meaning free list.
• Mutator and collector (marker and sweeper).
• Tricolor marking (whether by header bit or
  bitmap. Grey bit for mutator-marker cooperation).
• No black cell can point to a white cell holds
  for mutator and marker (P1)

6
Mark-During-Sweep - terminology (cont)

• Marker phase is finished when all Grey cell
  vanish.
• Sweeper claims all white cells as garbage.
• Two generations operation - sweeping current
  while marking the next.
• Working schemes and (P1) hold for generation of
  the marker.

7
Mark-During-Sweep - terminology (cont).

• Comparisons

8
The Algorithm - conventions.

• Left(i), right(i), colourN (i), shadeN(i).
• Cell is

9
The Algorithm - conventions (cont).

• ShadeN(i) if the cell is white turn it to Grey.

10
The Marker.
11
The Marker. (cont)

• Termination is ensured.
• Not the very efficient to simplify proof.

12
The sweeper.
13
The Sweeper. (cont)

• Shading is necessary for (P1).
• WhiteN is necessary for the N2 marker.

14
The collector.
15
The Mutator.

• Whenever an edge is redirected the target of the
  edge is shadeG1ed.
• G1 is the generation of the current marker.

16
Case Analysis.
17
Incremental GC .

• Memory size (M with 2M ptrs).
• Mutator consumption speed (vE cells to unit of
  times)
• Sweeper rate (rS ratio of reclaimed to swept
  cells in unit of time).
• Free list length (L, and is instantenous ).
• Marker rate (rM rate of marked/ visited).

18
Incremental GC. (cont)

• No mutator starvation.
• M/L cells to visit by the marker and the sweeper.
• L regulate the time of the GC.
• Two kind of real-time tasks can be identified.

19
Variants

• Flip flops rather than incrementing G
• Blackening instead of shading in the sweeper.
• Using a global counter of Grey cells instead of
  rescanning the heap (marker).
• Mark recursively as much as possible then collect
  left Greys.

20
Conclusions

• An algorithm for real-time embedded systems that
  has predictable response time, guaranteed
  throughput and can work with narrow memories.
• Mutator marker and sweeper run concurrently.

21
The END