Ravenbrook / Projects / Memory Pool System / Master Product Sources / Design Documents

Memory Pool System Project

Status

This document is currently a mixture of very old design notes (the preformatted section immediately following) and some newer stuff. It doesn't yet form anything like a complete picture. 

                                 TRACER
                            design.mps.trace
                           incomplete design
                             drj 1996-09-25


ARCHITECTURE:

.instance.limit: There will be a limit on the number of traces that can be 
created at any one time.  This effectively limits the number of concurrent 
traces.  This limitation is expressed in the symbol TRACE_MAX [currently set to 
1, see request.mps.160020 "Multiple traces would not work"  drj 1998-06-15].

.rate: [see mail.nickb.1997-07-31.14-37].  [Now revised?  See 
request.epcore.160062 and change.epcore.minnow.160062.  drj 1998-06-15]

.exact.legal: Exact references should either point outside the arena (to 
non-managed address space) or to a tract allocated to a pool.  Exact references 
that are to addresses which the arena has reserved but hasn't allocated memory 
to are illegal (the exact reference couldn't possibly refer to a real object).  
Depending on the future semantics of PoolDestroy we might need to adjust our 
strategy here.  See mail.dsm.1996-02-14.18-18 for a strategy of coping 
gracefully with PoolDestroy.  We check that this is the case in the fixer.  It 
may be sensible to make this check CRITICAL in certain configurations.

.fix.fixed.all: ss->fixedSummary is accumulated (in the fixer) for all the 
pointers whether or not they are genuine references.  We could accumulate fewer 
pointers here; if a pointer fails the TractOfAddr test then we know it isn't a 
reference, so we needn't accumulate it into the fixed summary.  The design 
allows this, but it breaks a useful post-condition on scanning (if the 
accumulation of ss->fixedSummary was moved the accuracy of ss->fixedSummary 
would vary according to the "width" of the white summary).  See 
mail.pekka.1998-02-04.16-48 for improvement suggestions.


ANALYSIS:

.fix.copy-fail: Fixing can always succeed, even if copying the referenced 
object has failed (due to lack of memory, for example), by backing off to 
treating a reference as ambiguous.  Assuming that fixing an ambiguous reference 
doesn't allocate memory (which is no longer true for AMC for example).  See 
request.dylan.170560 for a slightly more sophisticated way to proceed when you 
can no longer allocate memory for copying.


IDEAS:

.flip.after: To avoid excessive barrier impact on the mutator immediately after 
flip, we could scan during flip other objects which are "near" the roots, or 
otherwise known to be likely to be accessed in the near future.


IMPLEMENTATION:

Speed

.fix: The fix path is critical to garbage collection speed.  Abstractly fix is 
applied to all the references in the non-white heap and all the references in 
the copied heap.  Remembered sets cut down the number of segments we have to 
scan.  The zone test cuts down the number of references we call fix on.  The 
speed of the remainder of the fix path is still critical to system 
performance.  Various modifications to and aspects of the system are concerned 
with maintaining the speed along this path.

.fix.tractofaddr: TractOfAddr is called on every reference that passes the zone 
test and is on the critical path, to determine whether the segment is white. 
There is no need to examine the segment to perform this test, since whiteness 
information is duplicated in tracts, specifically to optimize this test.  
TractOfAddr itself is a simple class dispatch function (which dispatches to the 
arena class's TractOfAddr method).  Inlining the dispatch and inlining the 
functions called by VMTractOfAddr makes a small but noticable difference to the 
speed of the dylan compiler.

.fix.noaver: AVERs in the code add bulk to the code (reducing I-cache efficacy) 
and add branches to the path (polluting the branch pedictors) resulting in a 
slow down.  Removing all the AVERs from the fix path improves the overall speed 
of the dylan compiler by as much as 9%.

.fix.nocopy: AMCFix used to copy objects by using the format's copy method.  
This involved a function call (through an indirection) and in dylan_copy a call 
to dylan_skip (to recompute the length) and call to memcpy with general 
parameters.  Replacing this with a direct call to memcpy removes these 
overheads and the call to memcpy now has aligned parameters.  The call to 
memcpy is inlined by the (C) compiler.  This change results in a 4-5% speed-up 
in the dylan compiler.

.reclaim: Because the reclaim phase of the trace (implemented by TraceReclaim) 
examines every segment it is fairly time intensive.  rit's profiles presented 
in request.dylan.170551 show a gap between the two varieties variety.hi and 
variety.wi.

.reclaim.noaver: Converting AVERs in the loops of TraceReclaim, PoolReclaim, 
AMCReclaim (LOReclaim? AWLReclaim) will result in a noticeable speed 
improvement [insert actual speed improvement here].

Life cycle of a trace object

TraceCreate creates a trace in state TraceINIT

Some segments get condemned (made white).

TraceStart gets called which:

  • Derives an initial reference partition based on the existing white set. The white zone set and the segments' summaries are used to create an initial grey set.

  • emits a GCStart message.

  • initialises trace->rate by estimating the required scanning rate

  • moves the trace into the state TraceUNFLIPPED

  • immediately calls traceFlip which flips the trace and moves it into state TraceFLIPPED.

Whilst a trace is alive every so often its traceQuantum method gets invoked (via TracePoll) in order to do a quantum of tracing work. traceQuantum is responsible for ticking through the trace's top-level state machine. Most of the interesting work, the tracing, happens in the TraceFLIPPED state.

The trace transitions through its states in the following sequence: TraceINIT -> (TraceUNFLIPPED) -> TraceFLIPPED -> TraceRECLAIM -> TraceFINISHED.

Whilst TraceUNFLIPPED appears in the code, no trace does any work in this state; all traces are immediately flipped to be in the TraceFLIPPED state (see above).

Once the trace is in the TraceFINISHED state it performs no more work and it can be safely destroyed. Generally the callers of traceQuantum will destroy the trace.

Making Progress - Scanning Grey Segments

Most of the interesting work of a trace, the actual tracing, happens in the TraceFLIPPED state (work would happen in the TraceUNFLIPPED state, but that is not implemented).

The tracer makes progress by choosing a grey segment to scan, and scanning it. The actual scanning is performed by pools.

Note that at all times a reference partition is maintained.

The order in which the trace scans things determines the semantics of certain types of references (in particular, weak and final references). Or, to put it another way the desired semantics of weak and final references impose certain restrictions on the order in which the trace can scan things.

The tracer uses a system of reference ranks (or just ranks) so that it can impose an order on its scanning work. The ranks are ordered.

The tracer proceeds band by band. The first band is all objects it can reach by following references of the first rank. The second band is all subsequent objects it can reach by following references of the second and first ranks. The third band is all subsequent objects it can reach by following references of the third, second, and first ranks. And so on. The description of the tracer working like this originated in an e-mail message from RHSK [RHSK 2007-06-25].

A trace keep track of which band it is tracing. This is returned by the TraceBand method. Keeping this band information helps it implement the semantics of finalization and weakness. The band used to not be explicitly stored, but this hindered the implementation of good finalization semantics (essentially in some circumstances finalization messages were delayed by at least one collection cycle, see job001658).

The band is used when selecting a grey segment to scan (the selection occurs in traceFindGrey). The tracer attempts to first find segments whose rank is the current band, then segments whose rank is previous to the current band, and so on. If there are no segments found then the current band is exhausted and the current band is incremented to the next rank. When the current band is moved through all the ranks in this fashion there is no more tracing to be done.

A. References

[RHSK 2007-06-25] "the semantics of rank-based tracing"; RHSK; <URL: http://info.ravenbrook.com/mail/2007/06/25/11-35-57/0.txt>; 2007-06-25.

B. Document History

2002-06-07 RB Converted from MMInfo database design document.
2007-07-02 DRJ Added notes on tracer progress