MPS Strategy

author Nick Barnes
date 2013-06-04
index terms pair: strategy; design
organization Ravenbrook Limited
revision //info.ravenbrook.com/project/mps/master/design/strategy.txt#27
tag design.mps.strategy

Introduction

.intro This is the design of collection strategy for the MPS.

.readership MPS developers.

Overview

.overview The MPS uses "strategy" code to make three decisions:

  • when to start a collection trace;
  • what to condemn;
  • how to schedule tracing work.

This document describes the current strategy, identifies some weaknesses in it, and outlines some possible future development directions.

Requirements

[TODO: source some from req.dylan, or do an up-to-date requirements analysis -- NB 2013-03-25]

Garbage collection is a trade-off between time and space: it consumes some [CPU] time in order to save some [memory] space. Strategy shifts the balance point. A better strategy will take less time to produce more space. Examples of good strategy might include:

  • choosing segments to condemn which contain high proportions of dead objects;
  • starting a trace when a large number of objects have just died;
  • doing enough collection soon enough that the client program never suffers low-memory problems;
  • using otherwise-idle CPU resources for tracing.

Conversely, it would be bad strategy to do the reverse of each of these (condemning live objects; tracing when there's very little garbage; not collecting enough; tracing when the client program is busy).

Abstracting from these notions, requirements on strategy would relate to:

  • Maximum pause time and other utilization metrics (for example, bounded mutator utilization, minimum mutator utilization, total MPS CPU usage);
  • Collecting enough garbage (for example: overall heap size; low-memory requirements).
  • Allowing client control (for example, client recommendations for collection timing or condemnation).

There are other possible strategy considerations which are so far outside the scope of current strategy and MPS design that this document disregards them. For example, either inferring or allowing the client to specify preferred relative object locations ("this object should be kept in the same cache line as that one"), to improve cache locality.

Generations

The largest part of the current MPS strategy implementation is the support for generational garbage collections.

General data structures

The fundamental structure of generational garbage collection is the Chain, which describes a sequence of generations.

A chain specifies the "capacity" and "mortality" for each generation. When creating an automatically collected pool, the client code may specify the chain which will control collections for that pool. The same chain may be used for multiple pools. If no chain is specified, the pool uses the arena's default generation chain.

Each generation in a chain has a GenDesc structure, allocated in an array pointed to from the chain. In addition to the generations in the chains, the arena has a unique GenDesc structure, named topGen and described in comments as "the dynamic generation" (misleadingly: in fact it is the least dynamic generation).

Each automatically collected pool has a set of PoolGen structures, one for each generation that it can allocate or promote into. The PoolGen structures for each generation point to the GenDesc for that generation, and are linked together in a ring on the GenDesc. These structures are used to gather accounting information for strategy decisions.

The non-moving automatic pool classes (AMS, AWL and LO) do not support generational collection, so they allocate into a single generation. The moving automatic pool classes (AMC and AMCZ) have one pool generations for each generation in the chain, plus one pool generation for the arena's "top generation".

AMC data structures

An AMC pool creates an array of pool generation structures of type amcGen (a subclass of PoolGen). Each pool generation points to the forwarding buffer for that generation: this is the buffer that surviving objects are copied into.

AMC segments point to the AMC pool generation that the segment belongs to, and AMC buffers point to the AMC pool generation that the buffer will be allocating into.

The forwarding buffers are set up during AMC pool creation. Each generation forwards into the next higher generation in the chain, except for the top generation, which forwards to itself. Thus, objects are "promoted" up the chain of generations until they end up in the top generations, which is shared between all generational pools.

Collections

Collections in the MPS start in one of two ways:

  1. A collection of the world starts via TraceStartCollectAll(). This simply condemns all segments in all automatic pools.
  2. A collection of some set of generations starts via PolicyStartTrace(). See .policy.start.

Zones

Each generation in each chain has a zoneset associated with it (gen->zones); the condemned zoneset is the union of some number of generation's zonesets.

An attempt is made to use distinct zonesets for different generations. Segments in automatic pools are allocated using PoolGenAlloc() which creates a LocusPref using the zoneset from the generation's GenDesc. The zoneset for each generation starts out empty. If the zoneset is empty, an attempt is made to allocate from a free zone. The GenDesc zoneset is augmented with whichever zones the new segment occupies.

Note that this zoneset can never shrink.

Parameters

.param.intro: A generation has two parameters, capacity and mortality, specified by the client program.

.param.capacity: The capacity of a generation is the amount of new allocation in that generation (that is, allocation since the last time the generation was condemned) that will cause the generation to be collected by TracePoll().

.param.capacity.misnamed: The name capacity is unfortunate since it suggests that the total amount of memory in the generation will not exceed this value. But that will only be the case for pool classes that always promote survivors to another generation. When there is old allocation in the generation (that is, prior to the last time the generation was condemned), as there is in the case of non-moving pool classes, the size of a generation is unrelated to its capacity.

.param.mortality: The mortality of a generation is the proportion (between 0 and 1) of memory in the generation that is expected to be dead when the generation is collected. It is used in TraceStart() to estimate the amount of data that will have to be scanned in order to complete the trace.

Accounting

.accounting.intro: Pool generations maintain the sizes of various categories of data allocated in that generation for that pool. This accounting information is reported via the event system, but also used in two places:

.accounting.poll: ChainDeferral() uses the new size of each generation to determine which generations in the chain are over capacity and so might need to be collected by PolicyStartTrace().

.accounting.condemn: PolicyStartTrace() uses the new size of each generation to determine which generations in the chain will be collected; it also uses the total size of the generation to compute the mortality.

.accounting.check: Computing the new size for a pool generation is far from straightforward: see job003772 and job004007 for some (former) errors in this code. In order to assist with checking that this has been computed correctly, the locus module uses a double-entry book-keeping system to account for every byte in each pool generation. This uses seven accounts:

.account.total: Memory acquired from the arena.

.account.total.negated: From the point of view of the double-entry system, the total should be negative as it is owing to the arena, but it is inconvenient to represent negative sizes, and so the positive value is stored instead.

.account.total.negated.justification: We don't have a type for signed sizes; but if we represented it in two's complement using the unsigned Size type then Clang's unsigned integer overflow detector would complain.

.account.free: Memory that is not in use (free or lost to fragmentation).

.account.buffered: Memory in a buffer that was handed out to the client program via BufferFill(), and which has not yet been condemned.

.account.new: Memory in use by the client program, allocated since the last time the generation was condemned.

.account.old: Memory in use by the client program, allocated prior to the last time the generation was condemned.

.account.newDeferred: Memory in use by the client program, allocated since the last time the generation was condemned, but which should not cause collections via TracePoll(). (Due to ramping; see below.)

.account.oldDeferred: Memory in use by the client program, allocated prior to the last time the generation was condemned, but which should not cause collections via TracePoll(). (Due to ramping; see below.)

.accounting.op: The following operations are provided:

.accounting.op.alloc: Allocate a segment in a pool generation. Debit total, credit free. (But see .account.total.negated.)

.accounting.op.free: Free a segment. First, ensure that the contents of the segment are accounted as free, by artificially ageing any memory accounted as new or newDeferred (see .accounting.op.age) and then artifically reclaiming any memory accounted as old or oldDeferred (see .accounting.op.reclaim). Finally, debit free, credit total. (But see .account.total.negated.)

.accounting.op.fill: Fill a buffer. Debit free, credit buffered.

.accounting.op.empty: Empty a buffer. Debit buffered, credit new or newDeferred with the allocated part of the buffer, credit free with the unused part of the buffer.

.accounting.op.age: Condemn memory. Debit buffered (if part or all of a buffer was condemned) and either new or newDeferred, credit old or oldDeferred. Note that the condemned part of the buffer remains part of the buffer until the buffer is emptied, but is now accounted as old or oldDeferred. The uncondemned part of the buffer, if any, remains accounted as buffered until it is either emptied or condemned in its turn.

.accounting.op.reclaim: Reclaim dead memory. Debit old or oldDeferred, credit free.

.accounting.op.undefer: Stop deferring the accounting of memory. Debit oldDeferred, credit old. Debit newDeferred, credit new.

Ramps

The intended semantics of ramping are pretty simple. It allows the client to advise us of periods of large short-lived allocation on a particular AP. Stuff allocated using that AP during its "ramp" will probably be dead when the ramp finishes. How the MPS makes use of this advice is up to us, but for instance we might segregate those objects, collect them less enthusiastically during the ramp and then more enthusiastically soon after the ramp finishes. Ramps can nest.

A ramp is entered by calling:

mps_ap_alloc_pattern_begin(ap, mps_alloc_pattern_ramp())

or similar, and left in a similar way.

This is implemented on a per-pool basis, for AMC only (it's ignored by the other automatic pools). PoolAMC throws away the identity of the AP specified by the client. The implementation is intended to work by changing the generational forwarding behaviour, so that there is a "ramp generation" - one of the regular AMC generations - which forwards to itself if collected during a ramp (instead of promoting to an older generation). It also tweaks the strategy calculation code, in a way with consequences I am documenting elsewhere.

Right now, the code sets this ramp generation to the last generation specified in the pool's "chain": it ordinarily forwards to the "after-ramp" generation, which is the "dynamic generation" (i.e. the least dynamic generation, i.e. the arena-wide "top generation"). My recollection, and some mentions in design/poolamc, suggests that the ramp generation used to be chosen differently from this.

So far, it doesn't sound too ghastly, I guess, although the subversion of the generational system seems a little daft. Read on....

An AMC pool has a rampMode (which is really a state of a state machine), taking one of five values: OUTSIDE, BEGIN, RAMPING, FINISH, and COLLECTING (actually the enum values are called RampX for these X). We initialize in OUTSIDE. The pool also has a rampCount, which is the ramp nesting depth and is used to allow us to ignore ramp transitions other than the outermost. According to design/poolamc, there's an invariant (in BEGIN or RAMPING, rampCount > 0; in COLLECTING or OUTSIDE, rampCount == 0), but this isn't checked in AMCCheck() and in fact is false for COLLECTING (see below).

There is a small set of events causing state machine transitions:

  • entering an outermost ramp;
  • leaving an outermost ramp;
  • condemning any segment of a ramp generation (detected in AMCWhiten);
  • reclaiming any AMC segment.

Here's pseudo-code for all the transition events:

Entering an outermost ramp:
if not FINISH, go to BEGIN.
Leaving an outermost ramp:
if RAMPING, go to FINISH. Otherwise, go to OUTSIDE.
Condemning a ramp generation segment:
If BEGIN, go to RAMPING and make the ramp generation forward to itself (detach the forwarding buffer and reset its generation). If FINISH, go to COLLECTING and make the ramp generation forward to the after-ramp generation.
Reclaiming any AMC segment:
If COLLECTING:
if rampCount > 0, go to BEGIN. Otherwise go to OUTSIDE.

Now, some deductions:

  1. When OUTSIDE, the count is always zero, because (a) it starts that way, and the only ways to go OUTSIDE are (b) by leaving an outermost ramp (count goes to zero) or (c) by reclaiming when the count is zero.
  2. When BEGIN, the count is never zero (consider the transitions to BEGIN and the transition to zero).
  3. When RAMPING, the count is never zero (again consider transitions to RAMPING and the transition to zero).
  4. When FINISH, the count can be anything (the transition to FINISH has zero count, but the Enter transition when FINISH can change that and then it can increment to any value).
  5. When COLLECTING, the count can be anything (from the previous fact, and the transition to COLLECTING).
  6. This is a bug!! The ramp generation is not always reset (to forward to the after-ramp generation). If we get into FINISH and then see another ramp before the next condemnation of the ramp generation, we will Enter followed by Leave. The Enter will keep us in FINISH, and the Leave will take us back to OUTSIDE, skipping the transition to the COLLECTING state which is what resets the ramp generation forwarding buffer. [TODO: check whether I made an issue and/or fixed it; NB 2013-06-04]

The simplest change to fix this is to change the behaviour of the Leave transition, which should only take us OUTSIDE if we are in BEGIN or COLLECTING. We should also update design/poolamc to tell the truth, and check the invariants, which will be these:

OUTSIDE => zero BEGIN => non-zero RAMPING => non-zero

A cleverer change might radically rearrange the state machine (e.g. reduce the number of states to three) but that would require closer design thought and should probably be postponed until we have a clearer overall strategy plan.

While I'm writing pseudo-code versions of ramp-related code, I should mention this other snippet, which is the only other code relating to ramping (these notes are useful when thinking about the broader strategy code):

In AMCBufferFill(), if we're RAMPING, and filling the forwarding buffer of the ramp generation, and the ramp generation is the forwarding buffer's generation, set amcSeg->new to FALSE. Otherwise, add the segment size to poolGen.newSize.

And since I've now mentioned the amcSeg->new flag, here are the only other uses of that:

  • it initializes as TRUE.
  • When leaving an outermost ramp, go through all the segments in the pool. Any non-white segment in the rampGen with new set to FALSE has its size added to poolGen->newSize and gets new set to TRUE.
  • in amcSegWhiten(), if new is TRUE, the segment size is deducted from poolGen.newSize and new is set to FALSE.

Policy

.policy: Functions that make decisions about what action to take are collected into the policy module (policy.c). The purpose of doing so is to make it easier to understand this set of decisions and how they interact, and to make it easier to maintain and update the policy.

Assignment of zones

Res PolicyAlloc(Tract *tractReturn, Arena arena, LocusPref pref, Size size, Pool pool)

.policy.alloc: Allocate size bytes of memory on behalf of pool, based on the preferences described by pref. If successful, update *tractReturn to point to the first tract in the allocated memory and return ResOK. Otherwise, return a result code describing the problem, for example ResCOMMIT_LIMIT.

.policy.alloc.impl: This tries various methods in succession until one succeeds. First, it tries to allocate from the arena's free land in the requested zones. Second, it tries allocating from free zones. Third, it tries extending the arena and then trying the first two methods again. Fourth, it tries allocating from any zone that is not blacklisted. Fifth, it tries allocating from any zone at all.

.policy.alloc.issue: This plan performs poorly under stress. See for example job003898.

Deciding whether to collect the world

Bool PolicyShouldCollectWorld(Arena arena, double availableTime, Clock now, Clock clocks_per_sec)

.policy.world: Determine whether now is a good time for mps_arena_step() to start a collection of the world. Return TRUE if so, FALSE if not. The availableTime argument is an estimate of the time that's available for the collection, now is the current time as returned by ClockNow(), and clocks_per_sec is the result of calling ClocksPerSec().

.policy.world.impl: There are two conditions: the estimate of the available time must be enough to complete the collection, and the last collection of the world must be long enough in the past that the mps_arena_step() won't be spending more than a certain fraction of runtime in collections. (This fraction is given by the ARENA_MAX_COLLECT_FRACTION configuration parameter.)

Starting a trace

Bool PolicyStartTrace(Trace *traceReturn, Bool *collectWorldReturn, Arena arena, Bool collectWorldAllowed)

.policy.start: Consider starting a trace. If a trace was started, update *traceReturn to point to the trace and return TRUE. Otherwise, leave *traceReturn unchanged and return FALSE.

.policy.start.world: If collectWorldAllowed is TRUE, consider starting a collection of the whole world, and if such a collection is started, set *collectWorldReturn to TRUE.

This decision uses the "Lisp Machine" strategy, which tries to schedule collections of the world so that the collector just keeps pace with the mutator: that is, it starts a collection when the predicted completion time of the collection is around the time when the mutator is predicted to reach the current memory limit. See [Pirinen].

.policy.start.world.hack: The collectWorldAllowed flag was added to fix job004011 by ensuring that the MPS starts at most one collection of the world in each call to ArenaPoll(). But this is is fragile and inelegant. Ideally the MPS would be able to deduce that a collection of a set of generations can't possibly make progress (because nothing that refers to this set of generations has changed), and so not start such a collection.

.policy.start.chain: If collectWorldAllowed is FALSE, or if it is not yet time to schedule a collection of the world, PolicyStartTrace() considers collecting a set of zones corresponding to a set of generations on a chain.

It picks these generations by calling ChainDeferral() for each chain; this function indicates if the chain needs collecting, and if so, how urgent it is to collect that chain. The most urgent chain in need of collection (if any) is then condemned by calling policyCondemnChain(), which chooses the set of generations to condemn, and condemns all the segments in those generations.

Trace progress

Bool PolicyPoll(Arena arena)

.policy.poll: Return TRUE if the MPS should do some tracing work; FALSE if it should return to the mutator.

Bool PolicyPollAgain(Arena arena, Clock start, Bool moreWork, Work tracedWork)

.policy.poll.again: Return TRUE if the MPS should do another unit of work; FALSE if it should return to the mutator. start is the clock time when the MPS was entered; moreWork and tracedWork are the results of the last call to TracePoll().

.policy.poll.impl: The implementation keep doing work until either the maximum pause time is exceeded (see design.mps.arena.pause-time), or there is no more work to do. Then it schedules the next collection so that there is approximately one call to TracePoll() for every ArenaPollALLOCTIME bytes of allocation.

Document History

  • 2013-06-04 NB Checked this in although it's far from complete. Pasted in my 'ramping notes' from email, which mention some bugs which I may have fixed (TODO: check this).
  • 2014-01-29 RB The arena no longer manages generation zonesets.
  • 2014-05-17 GDR Bring data structures and condemn logic up to date.