preliminary notes on a nofl field-logging barrier

When you have a generational collector, you aim to trace only the part of the object graph that has been allocated recently. To do so, you need to keep a remembered set: a set of old-to-new edges, used as roots when performing a minor collection. A language run-time maintains this set by adding write barriers: little bits of collector code that run when a mutator writes to a field.

Whippet’s nofl space is a block-structured space that is appropriate for use as an old generation or as part of a sticky-mark-bit generational collector. It used to have a card-marking write barrier; see my article diving into V8’s new write barrier, for more background.

Unfortunately, when running whiffle benchmarks, I was seeing no improvement for generational configurations relative to whole-heap collection. Generational collection was doing fine in my tiny microbenchmarks that are part of Whippet itself, but when translated to larger programs (that aren’t yet proper macrobenchmarks), it was a lose.

I had planned on doing some serious tracing and instrumentation to figure out what was happening, and thereby correct the problem. I still plan on doing this, but instead for this issue I used the old noggin technique instead: just, you know, thinking about the thing, eventually concluding that unconditional card-marking barriers are inappropriate for sticky-mark-bit collectors. As I mentioned in the earlier article:

An unconditional card-marking barrier applies to stores to slots in all objects, not just those in oldspace; a store to a new object will mark a card, but that card may contain old objects which would then be re-scanned. Or consider a store to an old object in a more dense part of oldspace; scanning the card may incur more work than needed. It could also be that Whippet is being too aggressive at re-using blocks for new allocations, where it should be limiting itself to blocks that are very sparsely populated with old objects.

That’s three problems. The second is well-known. But the first and last are specific to sticky-mark-bit collectors, where pages mix old and new objects.

a precise field-logging write barrier

Back in 2019, Steve Blackburn’s paper Design and Analysis of Field-Logging Write Barriers took a look at the state of the art in precise barriers that record not regions of memory that have been updated, but the precise edges (fields) that were written to. He ends up re-using this work later in the 2022 LXR paper (see §3.4), where the write barrier is used for deferred reference counting and a snapshot-at-the-beginning (SATB) barrier for concurrent marking. All in all field-logging seems like an interesting strategy. Relative to card-marking, work during the pause is much less: you have a precise buffer of all fields that were written to, and you just iterate that, instead of iterating objects. Field-logging does impose some mutator cost, but perhaps the payoff is worth it.

To log each old-to-new edge precisely once, you need a bit per field indicating whether the field is logged already. Blackburn’s 2019 write barrier paper used bits in the object header, if the object was small enough, and otherwise bits before the object start. This requires some cooperation between the collector, the compiler, and the run-time that I wasn’t ready to pay for. The 2022 LXR paper was a bit vague on this topic, saying just that it used “a side table”.

In Whippet’s nofl space, we have a side table already, used for a number of purposes:

  1. Mark bits.

  2. Iterability / interior pointers: is there an object at a given address? If so, it will have a recognizable bit pattern.

  3. End of object, to be able to sweep without inspecting the object itself

  4. Pinning, allowing a mutator to prevent an object from being evacuated, for example because a hash code was computed from its address

  5. A hack to allow fully-conservative tracing to identify ephemerons at trace-time; this re-uses the pinning bit, since in practice such configurations never evacuate

  6. Bump-pointer allocation into holes: the mark byte table serves the purpose of Immix’s line mark byte table, but at finer granularity. Because of this though, it is swept lazily rather than eagerly.

  7. Generations. Young objects have a bit set that is cleared when they are promoted.

Well. Why not add another thing? The nofl space’s granule size is two words, so we can use two bits of the byte for field logging bits. If there is a write to a field, a barrier would first check that the object being written to is old, and then check the log bit for the field being written. The old check will be to a byte that is nearby or possibly the same as the one to check the field logging bit. If the bit is unsert, we call out to a slow path to actually record the field.

preliminary results

I disassembled the fast path as compiled by GCC and got something like this on x86-64, in AT&T syntax, for the young-generation test:

mov    %rax,%rdx
and    $0xffffffffffc00000,%rdx
shr    $0x4,%rax
and    $0x3ffff,%eax
or     %rdx,%rax
testb  $0xe,(%rax)

The first five instructions compute the location of the mark byte, from the address of the object (which is known to be in the nofl space). If it has any of the bits in 0xe set, then it’s in the old generation.

Then to test a field logging bit it’s a similar set of instructions. In one of my tests the data type looks like this:

struct Node {
  uintptr_t tag;
  struct Node *left;
  struct Node *right;
  int i, j;
};

Writing the left field will be in the same granule as the object itself, so we can just test the byte we fetched for the logging bit directly with testb against $0x80. For right, we should be able to know it’s in the same slab (aligned 4 MB region) and just add to the previously computed byte address, but the C compiler doesn’t know that right now and so recomputes. This would work better in a JIT. Anyway I think these bit-swizzling operations are just lost in the flow of memory accesses.

For the general case where you don’t statically know the offset of the field in the object, you have to compute which bit in the byte to test:

mov    %r13,%rcx
mov    $0x40,%eax
shr    $0x3,%rcx
and    $0x1,%ecx
shl    %cl,%eax
test   %al,%dil

Is it good? Well, it improves things for my whiffle benchmarks, relative to the card-marking barrier, seeing a 1.05×-1.5× speedup across a range of benchmarks. I suspect the main advantage is in avoiding the “unconditional” part of card marking, where a write to a new object could cause old objects to be added to the remembered set. There are still quite a few whiffle configurations in which the whole-heap collector outperforms the sticky-mark-bit generational collector, though; I hope to understand this a bit more by building a more classic semi-space nursery, and comparing performance to that.

Implementation links: the barrier fast-path, the slow path, and the sequential store buffers. (At some point I need to make it so that allocating edge buffers in the field set causes the nofl space to page out a corresponding amount of memory, so as to be honest when comparing GC performance at a fixed heap size.)

Until next time, onwards and upwards!

needed-bits optimizations in guile

Hey all, I had a fun bug this week and want to share it with you.

numbers and representations

First, though, some background. Guile’s numeric operations are defined over the complex numbers, not over e.g. a finite field of integers. This is generally great when writing an algorithm, because you don’t have to think about how the computer will actually represent the numbers you are working on.

In practice, Guile will represent a small exact integer as a fixnum, which is a machine word with a low-bit tag. If an integer doesn’t fit in a word (minus space for the tag), it is represented as a heap-allocated bignum. But sometimes the compiler can realize that e.g. the operands to a specific bitwise-and operation are within (say) the 64-bit range of unsigned integers, and so therefore we can use unboxed operations instead of the more generic functions that do run-time dispatch on the operand types, and which might perform heap allocation.

Unboxing is important for speed. It’s also tricky: under what circumstances can we do it? In the example above, there is information that flows from defs to uses: the operands of logand are known to be exact integers in a certain range and the operation itself is closed over its domain, so we can unbox.

But there is another case in which we can unbox, in which information flows backwards, from uses to defs: if we see (logand n #xff), we know:

  • the result will be in [0, 255]

  • that n will be an exact integer (or an exception will be thrown)

  • we are only interested in a subset of n‘s bits.

Together, these observations let us transform the more general logand to an unboxed operation, having first truncated n to a u64. And actually, the information can flow from use to def: if we know that n will be an exact integer but don’t know its range, we can transform the potentially heap-allocating computation that produces n to instead truncate its result to the u64 range where it is defined, instead of just truncating at the use; and potentially this information could travel farther up the dominator tree, to inputs of the operation that defines n, their inputs, and so on.

needed-bits: the |0 of scheme

Let’s say we have a numerical operation that produces an exact integer, but we don’t know the range. We could truncate the result to a u64 and use unboxed operations, if and only if only u64 bits are used. So we need to compute, for each variable in a program, what bits are needed from it.

I think this is generally known a needed-bits analysis, though both Google and my textbooks are failing me at the moment; perhaps this is because dynamic languages and flow analysis don’t get so much attention these days. Anyway, the analysis can be local (within a basic block), global (all blocks in a function), or interprocedural (larger than a function). Guile’s is global. Each CPS/SSA variable in the function starts as needing 0 bits. We then compute the fixpoint of visiting each term in the function; if a term causes a variable to flow out of the function, for example via return or call, the variable is recorded as needing all bits, as is also the case if the variable is an operand to some primcall that doesn’t have a specific needed-bits analyser.

Currently, only logand has a needed-bits analyser, and this is because sometimes you want to do modular arithmetic, for example in a hash function. Consider Bon Jenkins’ lookup3 string hash function:

#define rot(x,k) (((x)<<(k)) | ((x)>>(32-(k))))
#define mix(a,b,c) \
{ \
  a -= c;  a ^= rot(c, 4);  c += b; \
  b -= a;  b ^= rot(a, 6);  a += c; \
  c -= b;  c ^= rot(b, 8);  b += a; \
  a -= c;  a ^= rot(c,16);  c += b; \
  b -= a;  b ^= rot(a,19);  a += c; \
  c -= b;  c ^= rot(b, 4);  b += a; \
}
...

If we transcribe this to Scheme, we get something like:

(define (jenkins-lookup3-hashword2 str)
  (define (u32 x) (logand x #xffffFFFF))
  (define (shl x n) (u32 (ash x n)))
  (define (shr x n) (ash x (- n)))
  (define (rot x n) (logior (shl x n) (shr x (- 32 n))))
  (define (add x y) (u32 (+ x y)))
  (define (sub x y) (u32 (- x y)))
  (define (xor x y) (logxor x y))

  (define (mix a b c)
    (let* ((a (sub a c)) (a (xor a (rot c 4)))  (c (add c b))
           (b (sub b a)) (b (xor b (rot a 6)))  (a (add a c))
           (c (sub c b)) (c (xor c (rot b 8)))  (b (add b a))
           ...)
      ...))

  ...

These u32 calls are like the JavaScript |0 idiom, to tell the compiler that we really just want the low 32 bits of the number, as an integer. Guile’s compiler will propagate that information down to uses of the defined values but also back up the dominator tree, resulting in unboxed arithmetic for all of these operations.

(When writing this, I got all the way here and then realized I had already written quite a bit about this, almost a decade ago ago. Oh well, consider this your lucky day, you get two scoops of prose!)

the bug

All that was just prelude. So I said that needed-bits is a fixed-point flow analysis problem. In this case, I want to compute, for each variable, what bits are needed for its definition. Because of loops, we need to keep iterating until we have found the fixed point. We use a worklist to represent the conts we need to visit.

Visiting a cont may cause the program to require more bits from the variables that cont uses. Consider:

(define-significant-bits-handler
    ((logand/immediate label types out res) param a)
  (let ((sigbits (sigbits-intersect
                   (inferred-sigbits types label a)
                   param
                   (sigbits-ref out res))))
    (intmap-add out a sigbits sigbits-union)))

This is the sigbits (needed-bits) handler for logand when one of its operands (param) is a constant and the other (a) is variable. It adds an entry for a to the analysis out, which is an intmap from variable to a bitmask of needed bits, or #f for all bits. If a already has some computed sigbits, we add to that set via sigbits-union. The interesting point comes in the sigbits-intersect call: the bits that we will need from a are first the bits that we infer a to have, by forward type-and-range analysis; intersected with the bits from the immediate param; intersected with the needed bits from the result value res.

If the intmap-add call is idempotent—i.e., out already contains sigbits for a—then out is returned as-is. So we can check for a fixed-point by comparing out with the resulting analysis, via eq?. If they are not equal, we need to add the cont that defines a to the worklist.

The bug? The bug was that we were not enqueuing the def of a, but rather the predecessors of label. This works when there are no cycles, provided we visit the worklist in post-order; and regardless, it works for many other analyses in Guile where we compute, for each labelled cont (basic block), some set of facts about all other labels or about all other variables. In that case, enqueuing a predecessor on the worklist will cause all nodes up and to including the variable’s definition to be visited, because each step adds more information (relative to the analysis computed on the previous visit). But it doesn’t work for this case, because we aren’t computing a per-label analysis.

The solution was to rewrite that particular fixed-point to enqueue labels that define a variable (possibly multiple defs, because of joins and loop back-edges), instead of just the predecessors of the use.

Et voilà ! If you got this far, bravo. Type at y’all again soon!

fedi is for losers

Does the fediverse have a vibe? I think that yes, there’s a flave, and with reason: we have things in common. We all left Twitter, or refused to join in the first place. Many of us are technologists or tech-adjacent, but generally not startuppy. There is a pervasive do-it-yourself ethos. This last point often expresses itself as a reaction: if you don’t like it, then do it yourself, a different way. Make your own Mastoverse agent. Defederate. Switch instances. Fedi is the “patches welcome” of community: just fork it!

Fedi is freedom, in the sense of “feel free to send a patch”, which is also hacker-speak for “go fuck yourself”. We split; that’s our thing! Which, you know, no-platform the nazis and terfs, of course. It can be good and even necessary to cut ties with the bad. And yet, this is not a strategy for winning. What’s worse, it risks creating a feedback loop with losing, which is the topic of this screed.

alembics

Fedi distills losers: AI, covid, class war, climate, free software, on all of these issues, the sort of people that fedi attracts are those that lost. But, good news everyone: in fedi you don’t have to engage with the world, only with fellow losers! I know. I include myself in these sets. But beyond the fact that I don’t want to be a loser, it is imperative that we win: we can’t just give up on climate or class war. Thing is, we don’t have a plan to win, and the vibe I get from fedi is much more disengaged than strategic.

Twitter—and I admit, I loved Twitter, of yore—Twitter is now for the winners: the billionaires, the celebrities, the politicians. These attract the shills, the would-be’s. But winner is just another word for future has-been; nobody will gain power in the world via Twitter any more. Twitter continues to be a formidable force, but it wanes by the day.

Still, when I check my feed, there are some people I follow doing interesting work on Twitter: consider Tobi Haslett, Erin Pineda, Bree Newsome, Cédric Herrou, Louis Allday, Gabriel Winant, Hamilton Nolan, James Butler, Serge Slama: why there and not fedi? Sure, there is inertia: the network was woven on Twitter, not the mastoverse. But I am not sure that fedi is right for them, either. I don’t know that fedi is the kind of loser that is ready to get back in the ring and fight to win.

theories of power

What is fedi’s plan to win? If our model is so good, what are we doing to make it a dominant mode of social discourse, of using it as a vector to effect the changes we want to see in the world?

From where I sit, I don’t see that we have a strategy. Fedi is fine and all, but it doesn’t scare anyone. That’s not good enough. Twitter was always flawed but it was a great tool for activism and continues to be useful in some ways. Bluesky has some of that old-Twitter vibe, and perhaps it will supplant the original, in time; inshallah.

In the meantime, in fedi, I would like to suggest that with regards to the network itself, that we stop patting ourselves on the back. What we have is good but not good enough. We should aim to make the world a better place and neither complacency nor splitting are going to get us there. If fedi is to thrive, we need to get out of our own heads and make our community a place to be afraid of.

whippet progress update: feature-complete!

Greetings, gentle readers. Today, an update on recent progress in the Whippet embeddable garbage collection library.

feature-completeness

When I started working on Whippet, two and a half years ago already, I was aiming to make a new garbage collector for Guile. In the beginning I was just focussing on proving that it would be advantageous to switch, and also learning how to write a GC. I put off features like ephemerons and heap resizing until I was satisfied with the basics.

Well now is the time: with recent development, Whippet is finally feature-complete! Huzzah! Now it’s time to actually work on getting it into Guile. (I have some performance noodling to do before then, adding tracing support, and of course I have lots of ideas for different collectors to build, but I don’t have any missing features at the current time.)

heap resizing

When you benchmark a garbage collector (or a program with garbage collection), you generally want to fix the heap size. GC overhead goes down with more memory, generally speaking, so you don’t want to compare one collector at one size to another collector at another size.

(Unfortunately, many (most?) benchmarks of dynamic language run-times and the programs that run on them fall into this trap. Imagine you are benchmarking a program before and after a change. Automatic heap sizing is on. Before your change the heap size is 200 MB, but after it is 180 MB. The benchmark shows the change to regress performance. But did it really? It could be that at the same heap size, the change improved performance. You won’t know unless you fix the heap size.)

Anyway, Whippet now has an implementation of MemBalancer. After every GC, we measure the live size of the heap, and compute a new GC speed, as a constant factor to live heap size. Every second, in a background thread, we observe the global allocation rate. The heap size is then the live data size plus the square root of the live data size times a factor. The factor has two components. One is constant, the expansiveness of the heap: the higher it is, the more room the program has. The other depends on the program, and is computed as the square root of the ratio of allocation speed to collection speed.

With MemBalancer, the heap ends up getting resized at every GC, and via the heartbeat thread. During the GC it’s easy because it’s within the pause; no mutators are running. From the heartbeat thread, mutators are active: taking the heap lock prevents concurrent resizes, but mutators are still consuming empty blocks and producing full blocks. This works out fine in the same way that concurrent mutators is fine: shrinking takes blocks from the empty list one by one, atomically, and returns them to the OS. Expanding might reclaim paged-out blocks, or allocate new slabs of blocks.

However, even with some exponentially weighted averaging on the speed observations, I have a hard time understanding whether the algorithm is overall a good thing. I like the heartbeat thread, as it can reduce memory use of idle processes. The general square-root idea sounds nice enough. But adjusting the heap size at every GC feels like giving control of your stereo’s volume knob to a hyperactive squirrel.

GC collection time vs memory usage graph comparing V8 with MemBalancer. MemBalancer is shown to be better.
Figure 5 from the MemBalancer paper

Furthermore, the graphs in the MemBalancer paper are not clear to me: the paper claims more optimal memory use even in a single-heap configuration, but the x axis of the graphs is “average heap size”, which I understand to mean that maximum heap size could be higher than V8’s maximum heap size, taking into account more frequent heap size adjustments. Also, some measurement of total time would have been welcome, in addition to the “garbage collection time” on the paper’s y axis; there are cases where less pause time doesn’t necessarily correlate to better total times.

deferred page-out

Motivated by MemBalancer’s jittery squirrel, I implemented a little queue for use in paging blocks in and out, for the mmc and pcc collectors: blocks are quarantined for a second or two before being returned to the OS via madvise(MADV_DONTNEED). That way if you release a page and then need to reacquire it again, you can do so without bothering the kernel or other threads. Does it matter? It seems to improve things marginally and conventional wisdom says to not mess with the page table too much, but who knows.

mmc rename

Relatedly, Whippet used to be three things: the project itself, consisting of an API and a collection of collectors; one specific collector; and one specific space within that collector. Last time I mentioned that I renamed the whippet space to the nofl space. Now I finally got around to renaming what was the whippet collector as well: it is now the mostly-marking collector, or mmc. Be it known!

Also as a service note, I removed the “serial copying collector” (scc). It had the same performance as the parallel copying collector with parallelism=1, and pcc can be compiled with GC_PARALLEL=0 to explicitly choose the simpler serial grey-object worklist.

per-object pinning

The nofl space has always supported pinned objects, but it was never exposed in the API. Now it is!

Of course, collectors with always-copying spaces won’t be able to pin objects. If we want to support use of these collectors with embedders that require pinning, perhaps because of conservative root scanning, we’d need to switch to some kind of mostly-copying algorithm.

safepoints without allocation

Another missing feature was a safepoint API. It hasn’t been needed up to now because my benchmarks all allocate, but for programs that have long (tens of microseconds maybe) active periods without allocation, you want to be able to stop them without waiting too long. Well we have that exposed in the API now.

removed ragged stop

Following on my article on ragged stops, I removed ragged-stop marking from mmc, for a nice net 180 line reduction in some gnarly code. Speed seems to be similar.

next up: tracing

And with that, I’m relieved to move on to the next phase of Whippet development. Thanks again to NLnet for their support of this work. Next up, adding fine-grained tracing, so that I can noodle a bit on performance. Happy allocating!

conservative gc can be faster than precise gc

Should your garbage collector be precise or conservative? The prevailing wisdom is that precise is always better. Conservative GC can retain more objects than strictly necessary, making GC slow: GC has to more frequently, and it has to trace a larger heap on each collection. However the calculus is not as straightforward as most people think, and indeed there are some reasons to expect that conservative root-finding can result in faster systems.

(I have made / relayed some of these arguments before but I feel like a dedicated article can make a contribution here.)

problem precision

Let us assume that by conservative GC we mean conservative root-finding, in which the collector assumes that any integer on the stack that happens to be a heap address indicates a reference on the object containing that address. The address doesn’t have to be at the start of the object. Assume that objects on the heap are traced precisely; contrast to BDW-GC which generally traces both the stack and the heap conservatively. Assume a collector that will pin referents of conservative roots, but in which objects not referred to by a conservative root can be moved, as in Conservative Immix or Whippet’s stack-conservative-mmc collector.

With that out of the way, let’s look at some reasons why conservative GC might be faster than precise GC.

smaller lifetimes

A compiler that does precise root-finding will typically output a side-table indicating which slots in a stack frame hold references to heap objects. These lifetimes aren’t always precise, in the sense that although they precisely enumerate heap references, those heap references might actually not be used in the continuation of the stack frame. When GC occurs, it might mark more objects as live than are actually live, which is the imputed disadvantage of conservative collectors.

This is most obviously the case when you need to explicitly register roots with some kind of handle API: the handle will typically be kept live until the scope ends, but that might be an overapproximation of lifetime. A compiler that can assume conservative stack scanning may well exhibit more precision than it would if it needed to emit stack maps.

no run-time overhead

For generated code, stack maps are great. But if a compiler needs to call out to C++ or something, it needs to precisely track roots in a run-time data structure. This is overhead, and conservative collectors avoid it.

smaller stack frames

A compiler may partition spill space on a stack into a part that contains pointers to the heap and a part containing numbers or other unboxed data. This may lead to larger stack sizes than if you could just re-use a slot for two purposes, if the lifetimes don’t overlap. A similar concern applies for compilers that partition registers.

no stack maps

The need to emit stack maps is annoying for a compiler and makes binaries bigger. Of course it’s necessary for precise roots. But then there is additional overhead when tracing the stack: for each frame on the stack, you need to look up the stack map for the return continuation, which takes time. It may be faster to just test if words on the stack might be pointers to the heap.

unconstrained compiler

Having to make stack maps is a constraint imposed on the compiler. Maybe if you don’t need them, the compiler could do a better job, or you could use a different compiler entirely. A conservative compiler can sometimes have better codegen, for example by the use of interior pointers.

anecdotal evidence

The Conservative Immix paper shows that conservative stack scanning can beat precise scanning in some cases. I have reproduced these results with parallel-stack-conservative-mmc compared to parallel-mmc. It’s small—maybe a percent—but it was a surprising result to me and I thought others might want to know.

Also, Apple’s JavaScriptCore uses conservative stack scanning, and V8 is looking at switching to it. Funny, right?

conclusion

When it comes to designing a system with GC, don’t count out conservative stack scanning; the tradeoffs don’t obviously go one way or the other, and conservative scanning might be the right engineering choice for your system.

on taking advantage of ragged stops

Many years ago I read one of those Cliff Click “here’s what I learned” articles in which he was giving advice about garbage collector design, and one of the recommendations was that at a GC pause, running mutator threads should cooperate with the collector by identifying roots from their own stacks. You can read a similar assertion in their VEE2005 paper, The Pauseless GC Algorithm, though this wasn’t the source of the information.

One motivation for the idea was locality: a thread’s stack is already local to a thread. Then specifically in the context of a pauseless collector, you need to avoid races between the collector and the mutator for a thread’s stack, and having the thread visit its own stack neatly handles this problem.

However, I am not so interested any more in (so-called) pauseless collectors; though I have not measured myself, I am convinced enough by the arguments in the Distilling the real costs of production garbage collectors paper, which finds that state of the art pause-minimizing collectors actually increase both average and p99 latency, relative to a well-engineered collector with a pause phase. So, the racing argument is not so relevant to me, because a pause is happening anyway.

There was one more argument that I thought was interesting, which was that having threads visit their own stacks is a kind of cheap parallelism: the mutator threads are already there, they might as well do some work; it could be that it saves time, if other threads haven’t seen the safepoint yet. Mutators exhibit a ragged stop, in the sense that there is no clean cutoff time at which all mutators stop simultaneously, only a time after which no more mutators are running.

Visiting roots during a ragged stop introduces concurrency between the mutator and the collector, which is not exactly free; notably, it prevents objects marked while mutators are running from being evacuated. Still, it could be worth it in some cases.

Or so I thought! Let’s try to look at the problem analytically. Consider that you have a system with N processors, a stop-the-world GC with N tracing threads, and M mutator threads. Let’s assume that we want to minimize GC latency, as defined by the time between GC is triggered and the time that mutators resume. There will be one triggering thread that causes GC to begin, and then M–1 remote threads that need to reach a safepoint before the GC pause can begin.

The total amount of work that needs to be done during GC can be broken down into rootsi, the time needed to visit roots for mutator i, and then graph, the time to trace the transitive closure of live objects. We want to know whether it’s better to perform rootsi during the ragged stop or in the GC pause itself.

Let’s first look to the case where M is 1 (just one mutator thread). If we visit roots before the pause, we have

latencyragged,M=1 = roots0 + graphN

Which is to say, thread 0 triggers GC, visits its own roots, then enters the pause in which the whole graph is traced by all workers with maximum parallelism. It may be that graph tracing doesn’t fully parallelize, for example if the graph has a long singly-linked list, but the parallelism with be maximal within the pause as there are N workers tracing the graph.

If instead we visit roots within the pause, we have:

latencypause,M=1= roots0+graphN

This is strictly better than the ragged-visit latency.

If we have two threads, then we will need to add in some delay, corresponding to the time it takes for remote threads to reach a safepoint. Let’s assume that there is a maximum period (in terms of instructions) at which a mutator will check for safepoints. In that case the worst-case delay will be a constant, and we add it on to the latency. Let us assume also there are more than two threads available. The marking-roots-during-the-pause case it’s easiest to analyze:

latencypause,M=2= delay + roots0+roots1+graphN

In this case, a ragged visit could win: while the triggering thread is waiting for the remote thread to stop, it could perform roots0, moving the work out of the pause, reducing pause time, and reducing latency, for free.

latencyragged,M=2= delay + roots1 + graphN

However, we only have this win if the root-visiting time is smaller than the safepoint delay; otherwise we are just prolonging the pause. Thing is, you don’t know in general. If indeed the root-visiting time is short, relative to the delay, we can assume the roots elements of our equation are 0, and so the choice to mark during ragged stop doesn’t matter at all! If we assume instead that root-visiting time is long, then it is suboptimally parallelised: under-parallelised if we have more than M cores, oversubscribed if M is greater than N, and needlessly serializing before the pause while it waits for the last mutator to finish marking its roots. What’s worse, root-visiting actually slows down delay, because the oversubscribed threads compete with visitation for CPU time.

So in summary, I plan to switch away from doing GC work during the ragged stop. It is complexity that doesn’t pay. Onwards and upwards!

whippet update: faster evacuation, eager sweeping of empty blocks

Good evening. Tonight, notes on things I have learned recently while hacking on the Whippet GC library.

service update

For some time now, the name Whippet has referred to three things. Firstly, it is the project as a whole, consisting of an include-only garbage collection library containing a compile-time configurable choice of specific collector implementations. Also, it is the name of a specific Immix-derived collector. Finally, it is the name of a specific space within that collector, in which objects are mostly marked in place but can be evacuated if appropriate.

Well, naming being one of the two hard problems of computer science, I can only ask for forgiveness and understanding. I have started fixing this situation with the third component, renaming the whippet space to the nofl space. Nofl stands for no-free-list, indicating that it’s a (mostly) mark space but which does bump-pointer allocation instead of using freelists. Also, it stands for novel, in the sense that as far as I can tell, it is a design that hasn’t been tried yet.

unreliable evacuation

Following Immix, the nofl space has always had optimistic evacuation. It prefers to mark objects in place, but if fragmentation gets too high, it will try to defragment by evacuating sparse blocks into a small set of empty blocks reserved for this purpose. If the evacuation reserve fills up, nofl will dynamically switch to marking in place.

My previous implementation was a bit daft: some subset of blocks would get marked as being evacuation targets, and evacuation would allocate into those blocks in ascending address order. Parallel GC threads would share a single global atomically-updated evacuation allocation pointer. As you can imagine, this was a serialization bottleneck; I initially thought it wouldn’t be so important but for some workloads it is.

I had chosen this strategy to maximize utilization of the evacuation reserve; if you had 8 GC workers, each allocating into their own block, their blocks won’t all be full at the end of GC; that would waste space.

But reliability turns out to be unimportant. It’s more important to let parallel GC threads do their thing without synchronization, to the extent possible.

Also, this serialized allocation discipline imposed some complexity on the nofl space implementation; the evacuation allocator was not like the “normal” allocator. With worker-local allocation buffers, these two allocators are now essentially the same. (They differ in that the normal allocator interleaves lazy sweeping with allocation, and can allocate into blocks with survivors from the last GC instead of requiring empty blocks.)

eager sweeping

Another initial bad idea I had was to lean too much on lazy sweeping as a design principle. The idea was that deferring sweeping work until just before an allocator would write to a block would minimize cache overhead (you page in a block right when you will use it) and provide for workload-appropriate levels of parallelism (multiple mutator threads naturally parallelize sweeping).

Lazy sweeping was very annoying when it came to discovery of empty blocks. Empty blocks are precious: they can be returned to the OS if needed, they are useful for evacuation, and they have nice allocation properties, in that you can just do bump-pointer from beginning to end.

Nofl was discovering empty blocks just in time, from the allocator. If the allocator acquired a new block and found that it was empty, it would return it to a special list of empty blocks. Only if all sweepable pages were exhausted would an allocator use an empty block. But to prevent an allocator from pausing forever, there was a limit to the number of empty swept blocks that would be returned to the collector (10, as it happens); an 11th empty swept block would be given to a mutator for allocation. And so on and so on. Complicated, and you only know the number of empty blocks yielded by the last collection when the whole next allocation cycle has finished.

The obvious solution is some kind of separate mark on blocks, in addition to a mark on objects. I didn’t do it initially out of fear of overhead; marking is a fast path. The implementation I ended up making was a packed bitvector, with one bit per 64 kB block, at the beginning of each 4 MB slab of blocks. The beginning blocks are for metadata like this. For reasons, I don’t have the space for full bytes. When marking an object, if I see that a block’s mark is unset, I do an atomic_fetch_or_explicit on the byte with memory_order_relaxed ordering. In this way I only do the atomic op very rarely. It seems that on ARMv8.1 there is actually an instruction to do atomic bit setting; everywhere else it’s a garbage compare-and-swap thing, but on my x64 machine it’s fine.

Then after collection, during the pause, if I see a block is unmarked, I move it directly to the empty set instead of sweeping it. (I could probably do this concurrently outside the pause, but that would be for another day.)

And the overhead? Negative! Negative, in the sense that because I don’t have to sweep empty blocks, and that (for some workloads) collection can produce a significant-enough fraction of empty blocks, I actually see speedups with this strategy, relative to lazy sweeping. It also simplifies the allocator (no need for that return-the-11th-block logic).

The only wrinkle is as regards generational collection: nofl currently uses the sticky mark bit algorithm, which has to be applied also to block marks. Subtle, but not complicated.

fin

Next up is growing and shrinking the nofl-using Whippet collector (which might need another name), using the membalancer algorithm, and then I think I will be ready to move on to getting Whippet into Guile. Until then, happy hacking!

javascript weakmaps should be iterable

Good evening. Tonight, a brief position statement: it is a mistake for JavaScript’s WeakMap to not be iterable, and we should fix it.

story time

A WeakMap associates a key with a value, as long as the key is otherwise reachable in a program. (It is an ephemeron table.)

When WeakMap was added to JavaScript, back in the ES6 times, some implementors thought that it could be reasonable to implement weak maps not as a data structure in its own right, but rather as a kind of property on each object. Under the hood, adding an key→value association to a map M would set key[M] = value. GC would be free to notice dead maps and remove their associations in live objects.

If you implement weak maps like this, or are open to the idea of such an implementation, then you can’t rely on the associations being enumerable from the map itself, as they are instead spread out among all the key objects. So, ES6 specified WeakMap as not being iterable; given a map, you can’t know what’s in it.

As with many things GC-related, non-iterability of weak maps then gained a kind of legendary status: the lore states that non-iterability preserves some key flexibility for JS implementations, and therefore it is good, and you just have to accept it and move on.

dynamics

Time passed, and two things happened.

One was that this distributed WeakMap implementation strategy did not pan out; everyone ended up implementing weak maps as their own kind of object, and people use an algorithm like the one Steve Fink described a couple years ago to compute the map×key⇒value conjunction. The main original motivation for non-iterability was no longer valid.

The second development was WeakRef and FinalizationRegistry, which expose some details of reachability as viewed by the garbage collector to user JS code. With WeakRef (and WeakMap), you can build an iterable WeakMap.

(Full disclosure: I did work on ES6 and had a hand in FinalizationRegistry but don’t do JS language work currently.)

Thing is, your iterable WeakMap is strictly inferior to what the browser can provide: its implementation is extraordinarily gnarly, shipped over the wire instead of already in the browser, uses more memory, is single-threaded and high-latency (because FinalizationRegistry), and non-standard. What if instead as language engineers we just did our jobs and standardized iterability, as we do with literally every other collection in the JS standard?

Just this morning I wrote yet another iterable WeakSet (which has all the same concerns as WeakMap), and while it’s sufficient for my needs, it’s not good (lacking prompt cleanup of dead entries), and by construction can’t be great (because it has to be redundantly implemented on top of WeakSet instead of being there already).

I am sympathetic to deferring language specification decisions to allow the implementation space to be explored, but when the exploration is done and the dust has settled, we shouldn’t hesitate to pick a winner: JS weak maps and sets should be iterable. Godspeed, brave TC39 souls; should you take up this mantle, you are doing the Lord’s work!

Thanks to Philip Chimento for notes on the timeline and Joyee Cheung for notes on the iterable WeakMap implementation in the WeakRef spec. All errors mine, of course!

whippet progress update: funding, features, future

Greets greets! Today, an update on recent progress in Whippet, including sponsorship, a new collector, and a new feature.

the lob, the pitch

But first, a reminder of what the haps: Whippet is a garbage collector library. The target audience is language run-time authors, particularly “small” run-times: wasm2c, Guile, OCaml, and so on; to a first approximation, the kinds of projects that currently use the Boehm-Demers-Weiser collector.

The pitch is that if you use Whippet, you get a low-fuss small dependency to vendor into your source tree that offers you access to a choice of advanced garbage collectors: not just the conservative mark-sweep collector from BDW-GC, but also copying collectors, an Immix-derived collector, generational collectors, and so on. You can choose the GC that fits your problem domain, like Java people have done for many years. The Whippet API is designed to be a no-overhead abstraction that decouples your language run-time from the specific choice of GC.

I co-maintain Guile and will be working on integrating Whippet in the next months, and have high hopes for success.

bridgeroos!

I’m delighted to share that Whippet was granted financial support from the European Union via the NGI zero core fund, administered by the Dutch non-profit, NLnet foundation. See the NLnet project page for the overview.

This funding allows me to devote time to Whippet to bring it from proof-of-concept to production. I’ll finish the missing features, spend some time adding tracing support, measuring performance, and sanding off any rough edges, then work on integrating Whippet into Guile.

This bloggery is a first update of the progress of the funded NLnet project.

a new collector!

I landed a new collector a couple weeks ago, a parallel copying collector (PCC). It’s like a semi-space collector, in that it always evacuates objects (except large objects, which are managed in their own space). However instead of having a single global bump-pointer allocation region, it breaks the heap into 64-kB blocks. In this way it supports multiple mutator threads: mutators do local bump-pointer allocation into their own block, and when their block is full, they fetch another from the global store.

The block size is 64 kB, but really it’s 128 kB, because each block has two halves: the active region and the copy reserve. It’s a copying collector, after all. Dividing each block in two allows me to easily grow and shrink the heap while ensuring there is always enough reserve space.

Blocks are allocated in 64-MB aligned slabs, so you get 512 blocks in a slab. The first block in a slab is used by the collector itself, to keep metadata for the rest of the blocks, for example a chain pointer allowing blocks to be collected in lists, a saved allocation pointer for partially-filled blocks, whether the block is paged in or out, and so on.

The PCC not only supports parallel mutators, it can also trace in parallel. This mechanism works somewhat like allocation, in which multiple trace workers compete to evacuate objects into their local allocation buffers; when an allocation buffer is full, the trace worker grabs another, just like mutators do.

However, unlike the simple semi-space collector which uses a Cheney grey worklist, the PCC uses the fine-grained work-stealing parallel tracer originally developed for Whippet’s Immix-like collector. Each trace worker maintains a local queue of objects that need tracing, which currently has 1024 entries. If the local queue becomes full, the worker will publish 3/4 of those entries to the worker’s shared worklist. When a worker runs out of local work, it will first try to remove work from its own shared worklist, then will try to steal from other workers.

Of course, because threads compete to evacuate objects, we have to use atomic compare-and-swap instead of simple forwarding pointer updates; if you only have one mutator thread and are OK with just one tracing thread, you can avoid the ~30% performance penalty that atomic operations impose. The PCC generally starts to win over a semi-space collector when it can trace with 2 threads, and gets better with each thread you add.

I sometimes wonder whether the worklist should contain grey edges or grey objects. MMTk seems to do the former, and bundles edges into work packets, which are the unit of work sharing. I don’t know yet what is best and look forward to experimenting once I have better benchmarks.

Anyway, maintaining an external worklist is cheating in a way: unlike the Cheney worklist, this memory is not accounted for as part of the heap size. If you are targetting a microcontroller or something, probably you need to choose a different kind of collector. Fortunately, Whippet enables this kind of choice, as it contains a number of collector implementations.

What about benchmarks? Well, I’ll be doing those soon in a more rigorous way. For now I will just say that it seems to behave as expected and I am satisfied; it is useful to have a performance oracle against which to compare other collectors.

finalizers!

This week I landed support for finalizers!

Finalizers work in all the collectors: semi, pcc, whippet, and the BDW collector that is a shim to use BDW-GC behind the Whippet API. They have a well-defined relationship with ephemerons and are multi-priority, allowing embedders to build guardians or phantom references on top.

In the future I should do some more work to make finalizers support generations, if the collector is generational, allowing a minor collection to avoid visiting finalizers for old objects. But this is a straightforward extension that will happen at some point.

future!

And that’s the end of this update. Next up, I am finally going to tackle heap resizing, using the membalancer approach. Then basic Windows and Mac support, then I move on to the tracing and performance measurement phase. Onwards and upwards!

finalizers, guardians, phantom references, et cetera

Consider guardians. Compared to finalizers, in which the cleanup procedures are run concurrently with the mutator, by the garbage collector, guardians allow the mutator to control concurrency. See Guile’s documentation for more notes. Java’s PhantomReference / ReferenceQueue seems to be similar in spirit, though the details differ.

questions

If we want guardians, how should we implement them in Whippet? How do they relate to ephemerons and finalizers?

It would be a shame if guardians were primitive, as they are a relatively niche language feature. Guile has them, yes, but really what Guile has is bugs: because Guile implements guardians on top of BDW-GC’s finalizers (without topological ordering), all the other things that finalizers might do in Guile (e.g. closing file ports) might run at the same time as the objects protected by guardians. For the specific object being guarded, this isn’t so much of a problem, because when you put an object in the guardian, it arranges to prepend the guardian finalizer before any existing finalizer. But when a whole clique of objects becomes unreachable, objects referred to by the guarded object may be finalized. So the object you get back from the guardian might refer to, say, already-closed file ports.

The guardians-are-primitive solution is to add a special guardian pass to the collector that will identify unreachable guarded objects. In this way, the transitive closure of all guarded objects will be already visited by the time finalizables are computed, protecting them from finalization. This would be sufficient, but is it necessary?

answers?

Thinking more abstractly, perhaps we can solve this issue and others with a set of finalizer priorities: a collector could have, say, 10 finalizer priorities, and run the finalizer fixpoint once per priority, in order. If no finalizer is registered at a given priority, there is no overhead. A given embedder (Guile, say) could use priority 0 for guardians, priority 1 for “normal” finalizers, and ignore the other priorities. I think such a facility could also support other constructs, including Java’s phantom references, weak references, and reference queues, especially when combined with ephemerons.

Anyway, all this is a note for posterity. Are there other interesting mutator-exposed GC constructs that can’t be implemented with a combination of ephemerons and multi-priority finalizers? Do let me know!