wingolog

Usually in this space I like to share interesting things that I find out; you might call it a research-epistle-publish loop. Today, though, I come not with answers, but with questions, or rather one question, but with fractal surface area: what is the value proposition of generational garbage collection?

hypothesis

The conventional wisdom is encapsulated in a 2004 Blackburn, Cheng, and McKinley paper, “Myths and Realities: The Performance Impact of Garbage Collection”, which compares whole-heap mark-sweep and copying collectors to their generational counterparts, using the Jikes RVM as a test harness. (It also examines a generational reference-counting collector, which is an interesting predecessor to the 2022 LXR work by Zhao, Blackburn, and McKinley.)

The paper finds that generational collectors spend less time than their whole-heap counterparts for a given task. This is mainly due to less time spent collecting, because generational collectors avoid tracing/copying work for older objects that mostly stay in the same place in the live object graph.

The paper also notes an improvement for mutator time under generational GC, but only for the generational mark-sweep collector, which it attributes to the locality and allocation speed benefit of bump-pointer allocation in the nursery. However for copying collectors, generational GC tends to slow down the mutator, probably because of the write barrier, but in the end lower collector times still led to lower total times.

So, I expected generational collectors to always exhibit lower wall-clock times than whole-heap collectors.

test workbench

In whippet, I have a garbage collector with an abstract API that specializes at compile-time to the mutator’s object and root-set representation and to the collector’s allocator, write barrier, and other interfaces. I embed it in whiffle, a simple Scheme-to-C compiler that can run some small multi-threaded benchmarks, for example the classic Gabriel benchmarks. We can then test those benchmarks against different collectors, mutator (thread) counts, and heap sizes. I expect that the generational parallel copying collector takes less time than the whole-heap parallel copying collector.

results?

So, I ran some benchmarks. Take the splay-tree benchmark, derived from Octane’s splay.js. I have a port to Scheme, and the results are... not good!

In this graph the “pcc” series is the whole-heap copying collector, and “generational-pcc” is the generational counterpart, with a nursery sized such that after each collection, its size is 2 MB times the number of active mutator threads in the last collector. So, for this test with eight threads, on my 8-core Ryzen 7 7840U laptop, the nursery is 16MB including the copy reserve, which happens to be the same size as the L3 on this CPU. New objects are kept in the nursery one cycle before being promoted to the old generation.

There are also results for “mmc” and “generational-mmc” collectors, which use an Immix-derived algorithm that allows for bump-pointer allocation but which doesn’t require a copy reserve. There, the generational collectors use a sticky mark-bit algorithm, which has very different performance characteristics as promotion is in-place, and the nursery is as large as the available heap size.

The salient point is that at all heap sizes, and for these two very different configurations (mmc and pcc), generational collection takes more time than whole-heap collection. It’s not just the splay benchmark either; I see the same thing for the very different nboyer benchmark. What is the deal?

I am honestly quite perplexed by this state of affairs. I wish I had a narrative to tie this together, but in lieu of that, voici some propositions and observations.

“generational collection is good because bump-pointer allocation”

Sometimes people say that the reason generational collection is good is because you get bump-pointer allocation, which has better locality and allocation speed. This is misattribution: it’s bump-pointer allocators that have these benefits. You can have them in whole-heap copying collectors, or you can have them in whole-heap mark-compact or immix collectors that bump-pointer allocate into the holes. Or, true, you can have them in generational collectors with a copying nursery but a freelist-based mark-sweep allocator. But also you can have generational collectors without bump-pointer allocation, for free-list sticky-mark-bit collectors. To simplify this panorama to “generational collectors have good allocators” is incorrect.

“generational collection lowers pause times”

It’s true, generational GC does lower median pause times:

But because a major collection is usually slightly more work under generational GC than in a whole-heap system, because of e.g. the need to reset remembered sets, the maximum pauses are just as big and even a little bigger:

I am not even sure that it is meaningful to compare median pause times between generational and non-generational collectors, given that the former perform possibly orders of magnitude more collections than the latter.

Doing fewer whole-heap traces is good, though, and in the ideal case, the less frequent major traces under generational collectors allows time for concurrent tracing, which is the true mitigation for long pause times.

is it whiffle?

Could it be that the test harness I am using is in some way unrepresentative? I don’t have more than one test harness for Whippet yet. I will start work on a second Whippet embedder within the next few weeks, so perhaps we will have an answer there. Still, there is ample time spent in GC pauses in these benchmarks, so surely as a GC workload Whiffle has some utility.

One reasons that Whiffle might be unrepresentative is that it is an ahead-of-time compiler, whereas nursery addresses are assigned at run-time. Whippet exposes the necessary information to allow a just-in-time compiler to specialize write barriers, for example the inline check that the field being mutated is not in the nursery, and an AOT compiler can’t encode this as an immediate. But it seems a small detail.

Also, Whiffle doesn’t do much compiler-side work to elide write barriers. Could the cost of write barriers be over-represented in Whiffle, relative to a production language run-time?

Relatedly, Whiffle is just a baseline compiler. It does some partial evaluation but no CFG-level optimization, no contification, no nice closure conversion, no specialization, and so on: is it not representative because it is not an optimizing compiler?

is it something about the nursery size?

How big should the nursery be? I have no idea.

As a thought experiment, consider the case of a 1 kilobyte nursery. It is probably too small to allow the time for objects to die young, so the survival rate at each minor collection would be high. Above a certain survival rate, generational GC is probably a lose, because your program violates the weak generational hypothesis: it introduces a needless copy for all survivors, and a synchronization for each minor GC.

On the other hand, a 1 GB nursery is probably not great either. It is plenty large enough to allow objects to die young, but the number of survivor objects in a space that large is such that pause times would not be very low, which is one of the things you would like in generational GC. Also, you lose out on locality: a significant fraction of the objects you traverse are probably out of cache and might even incur TLB misses.

So there is probably a happy medium somewhere. My instinct is that for a copying nursery, you want to make it about as big as L3 cache, which on my 8-core laptop is 16 megabytes. Systems are different sizes though; in Whippet my current heuristic is to reserve 2 MB of nursery per core that was active in the previous cycle, so if only 4 threads are allocating, you would have a 8 MB nursery. Is this good? I don’t know.

is it something about the benchmarks?

I don’t have a very large set of benchmarks that run on Whiffle, and they might not be representative. I mean, they are microbenchmarks.

One question I had was about heap sizes. If a benchmark’s maximum heap size fits in L3, which is the case for some of them, then probably generational GC is a wash, because whole-heap collection stays in cache. When I am looking at benchmarks that evaluate generational GC, I make sure to choose those that exceed L3 size by a good factor, for example the 8-mutator splay benchmark in which minimum heap size peaks at 300 MB, or the 8-mutator nboyer-5 which peaks at 1.6 GB.

But then, should nursery size scale with total heap size? I don’t know!

Incidentally, the way that I scale these benchmarks to multiple mutators is a bit odd: they are serial benchmarks, and I just run some number of threads at a time, and scale the heap size accordingly, assuming that the minimum size when there are 4 threads is four times the minimum size when there is just one thread. However, multithreaded programs are unreliable, in the sense that there is no heap size under which they fail and above which they succeed; I quote:

"Consider 10 threads each of which has a local object graph that is usually 10 MB but briefly 100MB when calculating: usually when GC happens, total live object size is 10×10MB=100MB, but sometimes as much as 1 GB; there is a minimum heap size for which the program sometimes works, but also a minimum heap size at which it always works."

is it the write barrier?

A generational collector partitions objects into old and new sets, and a minor collection starts by visiting all old-to-new edges, called the “remembered set”. As the program runs, mutations to old objects might introduce new old-to-new edges. To maintain the remembered set in a generational collector, the mutator invokes write barriers: little bits of code that run when you mutate a field in an object. This is overhead relative to non-generational configurations, where the mutator doesn’t have to invoke collector code when it sets fields.

So, could it be that Whippet’s write barriers or remembered set are somehow so inefficient that my tests are unrepresentative of the state of the art?

I used to use card-marking barriers, but I started to suspect they cause too much overhead during minor GC and introduced too much cache contention. I switched to precise field-logging barriers some months back for Whippet’s Immix-derived space, and we use the same kind of barrier in the generational copying (pcc) collector. I think this is state of the art. I need to see if I can find a configuration that allows me to measure the overhead of these barriers, independently of other components of a generational collector.

is it something about the generational mechanism?

A few months ago, my only generational collector used the sticky mark-bit algorithm, which is an unconventional configuration: its nursery is not contiguous, non-moving, and can be as large as the heap. This is part of the reason that I implemented generational support for the parallel copying collector, to have a different and more conventional collector to compare against. But generational collection loses on some of these benchmarks in both places!

is it something about collecting more often?

On one benchmark which repeatedly constructs some trees and then verifies them, I was seeing terrible results for generational GC, which I realized were because of cooperative safepoints: generational GC collects more often, so it requires that all threads reach safepoints more often, and the non-allocating verification phase wasn’t emitting any safepoints. I had to change the compiler to emit safepoints at regular intervals (in my case, on function entry), and it sped up the generational collector by a significant amount.

This is one instance of a general observation, which is that any work that doesn’t depend on survivor size in a GC pause is more expensive with a generational collector, which runs more collections. Synchronization can be a cost. I had one bug in which tracing ephemerons did work proportional to the size of the whole heap, instead of the nursery; I had to specifically add generational support for the way Whippet deals with ephemerons during a collection to reduce this cost.

is it something about collection frequency?

Looking deeper at the data, I have partial answers for the splay benchmark, and they are annoying :)

Splay doesn’t actually allocate all that much garbage. At a 2.5x heap, the stock parallel MMC collector (in-place, sticky mark bit) collects... one time. That’s all. Same for the generational MMC collector, because the first collection is always major. So at 2.5x we would expect the generational collector to be slightly slower. The benchmark is simply not very good – or perhaps the most generous interpretation is that it represents tasks that allocate 40 MB or so of long-lived data and not much garbage on top.

Also at 2.5x heap, the whole-heap copying collector runs 9 times, and the generational copying collector does 293 minor collections and... 9 major collections. We are not reducing the number of major GCs. It means either the nursery is too small, so objects aren’t dying young when they could, or the benchmark itself doesn’t conform to the weak generational hypothesis.

At a 1.5x heap, the copying collector doesn’t have enough space to run. For MMC, the non-generational variant collects 7 times, and generational MMC times out. Timing out indicates a bug, I think. Annoying!

I tend to think that if I get results and there were fewer than, like, 5 major collections for a whole-heap collector, that indicates that the benchmark is probably inapplicable at that heap size, and I should somehow surface these anomalies in my analysis scripts.

collecting more often redux

Doing a similar exercise for nboyer at 2.5x heap with 8 threads (4GB for 1.6GB live data), I see that pcc did 20 major collections, whereas generational pcc lowered that to 8 major collections and 3471 minor collections. Could it be that there are still too many fixed costs associated with synchronizing for global stop-the-world minor collections? I am going to have to add some fine-grained tracing to find out.

conclusion?

I just don’t know! I want to believe that generational collection was an out-and-out win, but I haven’t yet been able to prove it is true.

I do have some homework to do. I need to find a way to test the overhead of my write barrier – probably using the MMC collector and making it only do major collections. I need to fix generational-mmc for splay and a 1.5x heap. And I need to do some fine-grained performance analysis for minor collections in large heaps.

Enough for today. Feedback / reactions very welcome. Thanks for reading and happy hacking!