Author(s): Michael Knyszek, Austin Clements
Last updated: 2019-10-18
The Go runtime's page allocator (i.e. (*mheap).alloc
) has scalability problems. In applications with a high rate of heap allocation and a high GOMAXPROCS, small regressions in the allocator can quickly become big problems.
Based on ideas from Austin about making P-specific land-grabs to reduce lock contention, and with evidence that most span allocations are one page in size and are for small objects (<=32 KiB in size), I propose we:
The Go runtime's page allocator (i.e. (*mheap).alloc
) has serious scalability issues. These were discovered when working through golang/go#28479 and kubernetes/kubernetes#75833 which were both filed during or after the Go 1.12 release. The common thread between each of these scenarios is a high rate of allocation and a high level of parallelism (in the Go world, a relatively high GOMAXPROCS value, such as 32).
As it turned out, adding some extra work for a small subset of allocations in Go 1.12 and removing a fast-path data structure in the page heap caused significant regressions in both throughput and tail latency. The fundamental issue is the heap lock: all operations in the page heap (mheap
) are protected by the heap lock (mheap.lock
). A high allocation rate combined with a high degree of parallelism leads to significant contention on this lock, even though page heap allocations are relatively cheap and infrequent. For instance, if the most popular allocation size is ~1 KiB, as seen with the Kubernetes scalability test, then the runtime accesses the page heap every 10th allocation or so.
The proof that this is really a scalability issue in the design and not an implementation bug in Go 1.12 is that we were seeing barging behavior on this lock in Go 1.11, which indicates that the heap lock was already in a collapsed state before the regressions in Go 1.12 were even introduced.
I believe we can significantly improve the scalability of the page allocator if we eliminate as much lock contention in the heap as possible. We can achieve this in two ways:
So, what is this common case? We currently have span allocation data for a couple large Go applications which reveal that an incredibly high fraction of allocations are for small object spans. First, we have data from Kubernetes' 12-hour load test, which indicates that 99.89% of all span allocations are for small object spans, with 93% being from the first 50 size classes, inclusive. Next, data from a large Google internal service shows that 95% of its span allocations are for small object spans, even though this application is known to make very large allocations relatively frequently. 94% of all of this application's span allocations are from the first 50 size classes, inclusive.
Thus, I propose we:
The goal is to have most (80%+) small object span allocations allocate quickly, and without a lock. (1) makes the allocator significantly more cache-friendly, predictable, and enables (2), which helps us avoid grabbing the lock in the common case and allows us to allocate a small number of pages very quickly.
Note that this proposal maintains the current first-fit allocation policy and highest-address-first scavenging policy.
With a first-fit policy, allocation of one page (the common case) amounts to finding the first free page in the heap. One promising idea here is to use a bitmap because modern microarchitectures are really good at iterating over bits. Each bit in the bitmap represents a single runtime page (8 KiB as of this writing), where 1 means in-use and 0 means free. “In-use” in the context of the new page allocator is now synonymous with “owned by a span”. The concept of a free span isn't useful here.
I propose that the bitmap be divided up into shards (called chunks) which are small enough to be quick to iterate over. 512-bit shards would each represent 4 MiB (8 KiB pages) and fit in roughly 2 cache lines. These shards could live in one global bitmap which is mapped in as needed, or to reduce the virtual memory footprint, we could use a sparse-array style structure like (*mheap).arenas
. Picking a chunk size which is independent of arena size simplifies the implementation because arena sizes are platform-dependent.
Simply iterating over the whole bitmap to find a free page is still fairly inefficient, especially for dense heaps. We want to be able to quickly skip over completely in-use sections of the heap. Thus, I propose we attach summary information to each chunk such that it‘s much faster to filter out chunks which couldn’t possibly satisfy the allocation.
What should this summary information contain? I propose we augment each chunk with three fields: start, max, end uintptr
. start
represents the number of contiguous 0 bits at the start of this bitmap shard. Similarly, end
represents the number of contiguous 0 bits at the end of the bitmap shard. Finally, max
represents the largest contiguous section of 0 bits in the bitmap shard.
The diagram below illustrates an example summary for a bitmap chunk. The arrow indicates which direction addresses go (lower to higher). The bitmap contains 3 zero bits at its lowest edge and 7 zero bits at its highest edge. Within the summary, there are 10 contiguous zero bits, which max
reflects.
With these three fields, we can determine whether we‘ll be able to find a sufficiently large contiguous free section of memory in a given arena or contiguous set of arenas with a simple state machine. Computing this summary information for an arena is less trivial to make fast, and effectively amounts to a combination of a table to get per-byte summaries and a state machine to merge them until we have a summary which represents the whole chunk. The state machine for start
and end
is mostly trivial. max
is only a little more complex: by knowing start
, max
, and end
for adjacent summaries, we can merge the summaries by picking the maximum of each summary’s max
value and the sum of their start
and end
values. I propose we update these summary values eagerly as spans are allocated and freed. For large allocations that span multiple arenas, we can zero out summary information very quickly, and we really only need to do the full computation of summary information for the ends of the allocation.
There‘s a problem in this design so far wherein subsequent allocations may end up treading the same path over and over. Unfortunately, this retreading behavior’s time complexity is O(heap * allocs)
. We propose a simple solution to this problem: maintain a hint address. A hint address represents an address before which there are definitely no free pages in the heap. There may not be free pages for some distance after it, hence why it is just a hint, but we know for a fact we can prune from the search everything before that address. In the steady-state, as we allocate from the lower addresses in the heap, we can bump the hint forward with every search, effectively eliminating the search space until new memory is freed. Most allocations are expected to allocate not far from the hint.
There's still an inherent performance problem with this design: larger allocations may require iterating over the whole heap, even with the hint address. This issue arises from the fact that we now have an allocation algorithm with a time complexity linear in the size of the heap. Modern microarchitectures are good, but not quite good enough to just go with this.
Therefore, I propose we take this notion of a summary-per-chunk and extend it: we can build a tree around this, wherein a given entry at some level of the radix tree represents the merge of some number of summaries in the next level. The leaf level in this case contains the per-chunk summaries, while each entry in the previous levels may reflect 8 chunks, and so on.
This tree would be constructed from a finite number of arrays of summaries, with lower layers being smaller in size than following layers, since each entry reflects a larger portion of the address space. More specifically, we avoid having an “explicit” pointer-based structure (think “implicit” vs. “explicit” when it comes to min-heap structures: the former tends to be an array, while the latter tends to be pointer-based).
Below is a diagram of the complete proposed structure.
The bottom two boxes are the arenas and summaries representing the full address space. Each red line represents a summary, and each set of dotted lines from a summary into the next layer reflects which part of that next layer that summary refers to.
In essence, because this tree reflects our address space, it is in fact a radix tree over addresses. By left shifting a memory address by different amounts, we can find the exact summary which contains that address in each level.
On allocation, this tree may be searched by looking at start
, max
, and end
at each level: if we see that max
is large enough, we continue searching in the next, more granular, level. If max
is too small, then we look to see if there‘s free space spanning two adjacent summaries’ memory regions by looking at the first‘s end
value and the second’s start
value. Larger allocations are therefore more likely to cross larger boundaries of the address space are more likely to get satisfied by levels in the tree which are closer to the root. Note that if the heap has been exhausted, then we will simply iterate over the root level, find all zeros, and return.
A number of details were omitted from the previous section for brevity, but these details are key for an efficient implementation.
Firstly, note that start, max, end uintptr
is an awkward structure in size, requiring either 12 bytes or 24 bytes to store naively, neither of which fits a small multiple of a cache line comfortably. To make this structure more cache-friendly, we can pack them tightly into 64-bits if we constrain the height of the radix tree. The packing is straight-forward: we may dedicate 21 bits to each of these three numbers and pack them into 63 bits. A small quirk with this scheme is that each of start
, max
, and end
are counts, and so we need to represent zero as well as the maximum value (2^21
), which at first glance requires an extra bit per field. Luckily, in that case (i.e. when max == 2^21
), then start == max && max == end
. We may use the last remaining bit to represent this case. A summary representing a completely full region is also conveniently uint64(0)
in this representation, which enables us to very quickly skip over parts of the address space we don't care about with just one load and branch.
As mentioned before, a consequence of this packing is that we need to place a restriction on our structure: each entry in the root level of the radix tree may only represent at most 2^21
8 KiB pages, or 16 GiB, because we cannot represent any more in a single summary. From this constraint, it follows that the root level will always be 2^(heapAddrBits - 21 - log2(pageSize))
in size in entries. Should we need to support much larger heaps, we may easily remedy this by representing a summary as start, max, end uint32
, though at the cost of cache line alignment and 1.5x metadata overhead. We may also consider packing the three values into two uint64
values, though at the cost of twice as much metadata overhead. Note that this concern is irrelevant on 32-bit architectures: we can easily represent the whole address space with a tree and 21 bits per summary field. Unfortunately, we cannot pack it more tightly on 32-bit architectures since at least 14 bits are required per summary field.
Now that we've limited the size of the root level, we need to pick the sizes of the subsequent levels. Each entry in the root level must reflect some number of entries in the following level, which gives us our fanout. In order to stay cache-friendly, I propose trying to keep the fanout close to the size of an L1 cache line or some multiple thereof. 64 bytes per line is generally a safe bet, and our summaries are 8 bytes wide, so that gives us a fanout of 8.
Taking all this into account, for a 48-bit address space (such as how we treat linux/amd64
in the runtime), I propose the following 5-level array structure:
16384
entries (fanout = 1, root)16384*8
entries (fanout = 8)16384*8*8
entries (fanout = 8)16384*8*8*8
entries (fanout = 8)16384*8*8*8*8
entries (fanout = 8, leaves)Note that level 4 has 2^48 bytes / (512 * 8 KiB)
entries, which is exactly the number of chunks in a 48-bit address space. Each entry at this level represents a single chunk. Similarly, since a chunk represents 512, or 2^9 pages, each entry in the root level summarizes a region of 2^21
contiguous pages, as intended. This scheme can be trivially applied to any system with a larger address space, since we just increase the size of the root level. For a 64-bit address space, the root level can get up to 8 GiB in size, but that‘s mostly virtual address space which is fairly cheap since we’ll only commit to what we use (see below).
For most heaps, 2^21
contiguous pages or 16 GiB per entry in the root level is good enough. If we limited ourselves to 8 entries in the root, we would still be able to gracefully support up to 128 GiB (and likely double that, thanks to prefetchers). Some Go applications may have larger heaps though, but as mentioned before we can always change the structure of a summary away from packing into 64 bits and then add an additional level to the tree, at the expense of some additional metadata overhead.
Overall this uses between KiB and hundreds of MiB of address space on systems with smaller address spaces (~600 MiB for a 48-bit address space, ~128 KiB for a 32-bit address space). For a full 64-bit address space, this layout requires ~37 TiB of reserved memory.
At first glance, this seems like an enormous amount, but in reality that‘s an extremely small fraction (~0.00022%) of the full address space. Furthermore, this address space is very cheap since we’ll only commit what we use, and to reduce the size of core dumps and eliminate issues with overcommit we will map the space as PROT_NONE
(only MEM_RESERVE
on Windows) and map it as read/write explicitly when we grow the heap (an infrequent operation).
There are only two known adverse effects of this large mapping on Linux:
ulimit -v
, which restricts even PROT_NONE
mappings.PROT_NONE
mappings. In the grand scheme of things, these are relatively minor consequences. The former is not used often, and in cases where it is, it‘s used as an inaccurate proxy for limiting a process’s physical memory use. The latter is mostly cosmetic, though perhaps some monitoring system uses it as a proxy for memory use, and will likely result in some harmless questions.I propose adding a free page cache to each P. The page cache, in essence, is a base address marking the beginning of a 64-page aligned chunk, and a 64-bit bitmap representing free pages in that chunk. With 8 KiB pages, this makes it so that at most each P can hold onto 512 KiB of memory.
The allocation algorithm would thus consist of a P first checking its own cache. If it‘s empty, it would then go into the bitmap and cache the first non-zero chunk of 64 bits it sees, noting the base address of those 64 bits. It then allocates out of its own cache if able. If it’s unable to satisfy the allocation from these bits, then it goes back and starts searching for contiguous bits, falling back on heap growth if it fails. If the allocation request is more than 16 pages in size, then we don't even bother checking the cache. The probability that N
consecutive free pages will be available in the page cache decreases exponentially as N
approaches 64, and 16 strikes a good balance between being opportunistic and being wasteful.
Note that allocating the first non-zero chunk of 64 bits is an equivalent operation to allocating one page out of the heap: fundamentally we're looking for the first free page we can find in both cases. This means that we can and should optimize for this case, since we expect that it will be extremely common. Note also that we can always update the hint address in this case, making all subsequent allocations (large and small) faster.
Finally, there‘s a little hiccup in doing this and that’s that acquiring an mspan
object currently requires acquiring the heap lock, since these objects are just taken out of a locked SLAB allocator. This means that even if we can perform the allocation uncontended we still need the heap lock to get one of these objects. We can solve this problem by adding a pool of mspan
objects to each P, similar to the sudog
cache.
With the elimination of free spans, scavenging must work a little differently as well. The primary bit of information we're concerned with here is the scavenged
field currently on each span. I propose we add a scavenged
bitmap to each heapArena
which mirrors the allocation bitmap, and represents whether that page has been scavenged or not. Allocating any pages would unconditionally clear these bits to avoid adding extra work to the allocation path.
The scavenger's job is now conceptually much simpler. It takes bits from both the allocation bitmap as well as the scavenged
bitmap and performs a bitwise-OR operation on the two to determine which pages are “scavengable”. It then scavenges any contiguous free pages it finds in a single syscall, marking the appropriate bits in the scavenged
bitmap. Like the allocator, it would have a hint address to avoid walking over the same parts of the heap repeatedly.
Because this new algorithm effectively requires iterating over the heap backwards, there‘s a slight concern with how much time it could take, specifically if it does the scavenge operation with the heap lock held like today. Instead, I propose that the scavenger iterate over the heap without the lock, checking the free and scavenged bitmaps optimistically. If it finds what appears to be valid set of contiguous scavengable bits, it’ll acquire the heap lock, verify their validity, and scavenge.
We're still scavenging with the heap lock held as before, but scaling the scavenger is outside the scope of this document (though we certainly have ideas there).
Another piece of the scavenging puzzle is how to deal with the fact that the current scavenging policy is huge-page aware. There are two dimensions to this huge-page awareness: the runtime counts the number of free and unscavenged huge pages for pacing purposes, and the runtime scavenges those huge pages first.
For the first part, the scavenger currently uses an explicit ratio calculated whenever the GC trigger is updated to determine the rate at which it should scavenge, and it uses the number of free and unscavenged huge pages to determine this ratio.
Instead, I propose that the scavenger releases memory one page at a time while avoiding breaking huge pages, and it times how long releasing each page takes. Given a 1% maximum time spent scavenging for the background scavenger, we may then determine the amount of time to sleep, thus effectively letting the scavenger set its own rate. In some ways this self-pacing is more accurate because we no longer have to make order-of-magnitude assumptions about how long it takes to scavenge. Also, it represents a significant simplification of the scavenger from an engineering perspective; there's much less state we need to keep around in general.
The downside to this self-pacing idea is that we must measure time spent sleeping and time spent scavenging, which may be funky in the face of OS-related context switches and other external anomalies (e.g. someone puts their laptop in sleep mode). We can deal with such anomalies by setting bounds on how high or low our measurements are allowed to go. Furthermore, I propose we manage an EWMA which we feed into the time spent sleeping to account for scheduling overheads and try to drive the actual time spent scavenging to 1% of the time the goroutine is awake (the same pace as before).
As far as scavenging huge pages first goes, I propose we just ignore this aspect of the current scavenger simplicity‘s sake. In the original scavenging proposal, the purpose of scavenging huge pages first was for throughput: we would get the biggest bang for our buck as soon as possible, so huge pages don’t get “stuck” behind small pages. There's a question as to whether this actually matters in practice: conventional wisdom suggests a first-fit policy tends to cause large free fragments to congregate at higher addresses. By analyzing and simulating scavenging over samples of real Go heaps, I think this wisdom mostly holds true.
The graphs below show a simulation of scavenging these heaps using both policies, counting how much of the free heap is scavenged at each moment in time. Ignore the simulated time; the trend is more important.
With the exception of two applications, the rest all seem to have their free and unscavenged huge pages at higher addresses, so the simpler policy leads to a similar rate of releasing memory. The simulation is based on heap snapshots at the end of program execution, so it's a little non-representative since large, long-lived objects, or clusters of objects, could have gotten freed just before measurement. This misrepresentation actually acts in our favor, however, since it suggests an even smaller frequency of huge pages appearing in the middle of the heap.
The purpose of this proposal is to help the memory allocator scale. To reiterate: it‘s current very easy to put the heap lock in a collapsing state. Every page-level allocation must acquire the heap lock, and with 1 KiB objects we’re already hitting that page on every 10th allocation.
To give you an idea of what kinds of timings are involved with page-level allocations, I took a trace from a 12-hour load test from Kubernetes when I was diagnosing kubernetes/kubernetes#75833. 92% of all span allocations were for the first 50 size classes (i.e. up and including 8 KiB objects). Each of those, on average, spent 4.0µs in the critical section with the heap locked, minus any time spent scavenging. The mode of this latency was between 3 and 4µs, with the runner-up being between 2 and 3µs. These numbers were taken with the load test built using Go 1.12.4 and from a linux/amd64
GCE instance. Note that these numbers do not include the time it takes to acquire or release the heap lock; it is only the time in the critical section.
I implemented a prototype of this proposal which lives outside of the Go runtime, and optimized it over the course of a few days. I then took heap samples from large, end-to-end benchmarks to get realistic heap layouts for benchmarking the prototype.
The prototype benchmark then started with these heap samples and allocated out of them until the heap was exhausted. Without the P cache, allocations took only about 680 ns on average on a similar GCE instance to the Kubernetes case, pretty much regardless of heap size. This number scaled gracefully relative to allocation size as well. To be totally clear, this time includes finding space, marking the space and updating summaries. It does not include clearing scavenge bits.
With the P cache included, that number dropped to 20 ns on average. The comparison with the P cache isn't an apples-to-apples comparison since it should include heap lock/unlock time on the slow path (and the k8s numbers should too). However I believe this only strengthens our case: with the P cache, in theory, the lock will be acquired less frequently, so an apples-to-apples comparison would be even more favorable to the P cache.
All of this doesn't even include the cost savings when freeing memory. While I do not have numbers regarding the cost of freeing, I do know that the free case in the current implementation is a significant source of lock contention (golang/go#28479). Each free currently requires a treap insertion and maybe one or two removals for coalescing.
In comparison, freeing memory in this new allocator is faster than allocation (without the cache): we know exactly which bits in the bitmap to clear from the address, and can quickly index into the arenas array to update them as well as their summaries. While updating the summaries still takes time, we can do even better by freeing many pages within the same arena at once, amortizing the cost of this update. In fact, the fast page sweeper that Austin added in Go 1.12 already iterates over the heap from lowest to highest address, freeing completely free spans. It would be straight-forward to batch free operations within the same heap arena to achieve this cost amortization.
In sum, this new page allocator design has the potential to not only solve our immediate scalability problem, but also gives us more headroom for future optimizations compared to the current treap-based allocator, for which a number of various caching strategies, have been designed and/or attempted.
The biggest way fragmentation could worsen with this design is as a result of the P cache. The P cache makes it so that allocation isn't quite exactly a serialized single-threaded first-fit, and P may hold onto pages which another P may need more.
In practice, given an in-tree prototype, we‘ve seen that this fragmentation scales with the number of P’s, and we believe this to be a reasonable trade-off: more processors generally require more memory to take advantage of parallelism.
As far as prior art is concerned, there hasn‘t been much work with bitmap allocators for languages which have a GC. Consider Go against other other managed languages with a GC: Go’s GC sits in a fairly unique point in the design space for GCs because it is a non-moving collector. Most allocators in other managed languages (e.g. Java) tend to be bump allocators, since they tend to have moving and compacting GCs. Other managed languages also tend to have many more allocations of smaller size compared to Go, so the “slow path” of allocating pages is usually just grabbing a fixed-size block out of a shared pool, which can be made to be quite fast. Go relies on being able to allocate blocks of different sizes to reduce fragmentation in its current allocator design.
However, when considering non-GC'd languages, e.g C/C++, there has been some notable work using bitmaps in memory allocators. In particular, DieHard, DieHarder, and Mesh. DieHard and DieHarder in particular implement an effecient amortized O(1)
bitmap-based allocator, though that was not their primary contribution. Mesh uses bitmaps for managing which slots are free in a span, like the Go allocator. We were not aware of DieHard(er) at the time of writing this proposal, though they use bitmaps to track individual objects instead of pages. There are also a few niche cases where they are used such as GPU-accelerated allocation and real-time applications.
A good point of comparison for Go‘s current page allocator is TCMalloc, and in many ways Go’s memory allocator is based on TCMalloc. However, there are some key differences that arise as a result of Go‘s GC. Notably, TCMalloc manages its per-CPU caches as arrays of pointers, rather than through spans directly like Go does. The reason for this, as far as I can tell, is because when a free occurs in TCMalloc, that object is immediately available for re-use, whereas with Go, object lifetimes are effectively rounded up to a GC cycle. As a result of this global, bulk (de)allocation behavior resulting in the lack of short-term re-use, I suspect Go tends to ask the page allocator for memory more often that TCMalloc does. This bulk (de)allocation behavior would thus help explain why page allocator scalability hasn’t been such a big issue for TCMalloc (again, as far as I'm aware).
In sum, Go sits in a unique point in the memory management design space. The bitmap allocator fits this point in the design space well: bulk allocations and frees can be grouped together to amortize the cost of updating the summaries thanks to the GC. Furthermore, since we don't move objects in the heap, we retain the flexibility of dealing with fragments efficiently through the radix tree.
One considered alternative is to keep the current span structure, and instead try to cache the spans themselves on a P, splitting them on each allocation without acquiring the heap lock.
While this seems like a good idea in principle, one big limitation is that you can only cache contiguous regions of free memory. Suppose many heap fragments tend to just be one page in size: one ends up having to go back to the page allocator every single time anyway. While it is true that one might only cache one page from the heap in the proposed design, this case is fairly rare in practice, since it picks up any available memory it can find in a given 64-page aligned region.
The proposed design also tends to have nicer properties: the treap structure scales logarithmically (probabilistically) with respect to the number of free heap fragments, but even this property doesn't scale too well to very large heaps; one might have to chase down 20 pointers in a 20 GiB heap just for an allocation, not to mention the additional removals required. Small heaps may see allocations as fast as 100 ns, whereas large heaps may see page allocation latencies of 4 µs or more on the same hardware. On the other hand, the proposed design has a very consistent performance profile since the radix tree is effectively perfectly balanced.
Furthermore, this idea of caching spans only helps the allocation case. In most cases the source of contention is not only allocation but also freeing, since we always have to do a treap insertion (and maybe one or two removals) on the free path. In this proposal, the free path is much more efficient in general (no complex operations required, just clearing memory), even though it still requires acquiring the heap lock.
Finally, caching spans doesn't really offer much headroom in terms of future optimization, whereas switching to a bitmap allocator allows us to make a variety of additional optimizations because the design space is mostly unexplored.
This proposal changes no public APIs in either syntax or semantics, and is therefore Go 1 backwards compatible.
Michael Knyszek will implement this proposal.
The implementation will proceed as follows:
mspan
objects for each P.