K-way merge with Go 1.22 range functions
This package is intended to be used as a library and installed with:
go get github.com/achille-roussel/kway-go
To enable the rangefunc experiment, set the GOEXPERIMENT environment variable in the shell that executes the go commands:
export GOEXPERIMENT=rangefunc
The package contains variations of the K-way merge algorithm for different forms of iterator sequences:
-
Merge and MergeFunc operate on sequences that yield single values. Merge must be used on ordered values, while MergeFunc accepts a comparison function as first argument to customize the ordering logic.
-
MergeSlice and MergeSliceFunc are similar functions but operate on sequences that yield slices of values. These are intended for applications with higher throughput requirements that use batching or read values from paging APIs.
The sequences being merged must each be ordered using the same comparison logic than the one used for the merge, or the algorithm will not be able to produce an ordered sequence of values.
The following code snippets illustrates how to merge three ordered sequences into one:
for v, err := range kway.Merge(seq0, seq1, seq2) {
...
}
More examples are available in the Go doc.
The merge functions report errors seen from the input sequences, but the presence of errors does not interrupt the merge operations. When an error occurs, it is immediately bubbled up to the program, but if more values are available in the input sequences, the program can continue consuming them after handling the error. This model delegates the decision of how to handle errors to the application, allowing it to carry or abort depending on the error value or type, for example:
for v, err := range kway.Merge(sequences...) {
if err != nil {
// handle the error, the program may choose to break out of the loop
// or carry on to read the next value.
...
} else {
// a value is available, process it
...
}
}
The K-way merge algorithm was inspired by the talk from Bryan Boreham at Gophercon 2023, which described how using a loser-tree instead of a min-heap improved performance of Loki's merge of log records.
The kway-go
package also adds a specialization for cases where the program
is merging exactly two sequences, since this can be implemented as a simple
union of two sets which has a much lower compute and memory footprint.
K-way merge is often used in stream processing or database engines to merge distributed query results into a single ordered result set. In those applications, performance of the underlying algorithms tend to matter: for example, when performing compaction of sorted records, the merge algorithm is on the critical path and often where most of the compute is being spent. In that regard, there are efficiency requirements that the implementation must fulfil to be a useful solution to those problems.
💡 While exploring the performance characteristics of the algorithm, it is important to keep in mind that absolute numbers are only useful in the context where they were collected, since measurements depend on the hardware executing the code, and the data being processed. We should use relative performance of different benchmarks within a given context as a hint to find opportunities for optimizations in production applications, not as universal truths.
The current implementation has already been optimized to maximize throughput, by amortizing as much of the baseline costs as possible, and ensure that CPU time is spent on the important parts of the algorithm.
As part of this optimization work, it became apparent that while the Go runtime
implementation of coroutines underneath iter.Pull2
has a much lower compute
footprint than using channels, it still has a significant overhead when reading
values in tight loops of the merge algorithm.
This graph shows a preview of the results, the full analysis is described in the following sections:
To explore performance, let's first establish a baseline. We use the throughput of merging a single sequence, which is simple reading all the values it yields as comparison point:
Merge1 592898557 1.843 ns/op 0 comp/op 542741115 merge/s
This benchmark shows that on this test machine, the highest theoretical throughput we can achieve is ~540M merge/s for one sequence, ~270M merge/s when merging two sequences, etc...
Now comparing the performance of merging two and three sequences:
Merge2 47742177 24.78 ns/op 0.8125 comp/op 40359389 merge/s
Merge3 27540648 42.23 ns/op 1.864 comp/op 23682342 merge/s
We observe a significant drop in throughput in comparison with iterating over a single sequence, with the benchmark now performing ~7x slower than the theoretical throughput limit.
The K-way merge algorithm has a complexity of O(n∙log(k)), there would also be a baseline cost for the added code implementing the merge operations, but almost an order of magnitude difference seems unexpected.
To understand what is happening, we can look into a CPU profile:
Duration: 3.46s, Total samples = 2.44s (70.45%)
Showing nodes accounting for 2.40s, 98.36% of 2.44s total
Dropped 9 nodes (cum <= 0.01s)
flat flat% sum% cum cum%
0.30s 12.30% 12.30% 0.72s 29.51% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].merge2[go.shape.int].func3
0.25s 10.25% 22.54% 0.34s 13.93% github.com/achille-roussel/kway-go.(*tree[go.shape.int]).next
0.21s 8.61% 31.15% 0.76s 31.15% github.com/achille-roussel/kway-go.sequence.func1
0.17s 6.97% 38.11% 0.26s 10.66% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].unbuffer[go.shape.int].func6.1
0.15s 6.15% 44.26% 0.25s 10.25% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func1.1
0.15s 6.15% 50.41% 0.21s 8.61% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func4.1
0.14s 5.74% 56.15% 0.23s 9.43% iter.Pull2[go.shape.[]go.shape.int,go.shape.interface { Error string }].func2
0.13s 5.33% 61.48% 0.13s 5.33% runtime/internal/atomic.(*Uint32).CompareAndSwap (inline)
0.11s 4.51% 65.98% 0.18s 7.38% iter.Pull2[go.shape.[]go.shape.int,go.shape.interface { Error string }].func1.1
0.10s 4.10% 70.08% 0.27s 11.07% runtime.coroswitch_m
0.09s 3.69% 73.77% 0.09s 3.69% github.com/achille-roussel/kway-go.benchmark[go.shape.int].func2
0.09s 3.69% 77.46% 0.09s 3.69% runtime.coroswitch
0.08s 3.28% 80.74% 0.11s 4.51% gogo
0.07s 2.87% 83.61% 0.09s 3.69% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func2.1
0.06s 2.46% 86.07% 0.06s 2.46% runtime.mapaccess1_fast64
0.05s 2.05% 88.11% 0.09s 3.69% github.com/achille-roussel/kway-go.benchmark[go.shape.int].func1
0.04s 1.64% 89.75% 0.04s 1.64% cmp.Compare[go.shape.int] (inline)
0.04s 1.64% 91.39% 0.04s 1.64% internal/race.Acquire
0.04s 1.64% 93.03% 0.04s 1.64% runtime.(*guintptr).cas (inline)
0.04s 1.64% 94.67% 0.32s 13.11% runtime.mcall
0.04s 1.64% 96.31% 0.04s 1.64% runtime.save_g
0.02s 0.82% 97.13% 0.02s 0.82% internal/race.Release
0.01s 0.41% 97.54% 0.43s 17.62% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].merge[go.shape.int].func5
0.01s 0.41% 97.95% 0.04s 1.64% github.com/achille-roussel/kway-go.nextNonEmptyValues[go.shape.int]
0.01s 0.41% 98.36% 0.08s 3.28% runtime/pprof.(*profMap).lookup
0 0% 98.36% 0.72s 29.51% github.com/achille-roussel/kway-go.BenchmarkMerge2
0 0% 98.36% 0.43s 17.62% github.com/achille-roussel/kway-go.BenchmarkMerge3
0 0% 98.36% 0.35s 14.34% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func1
0 0% 98.36% 0.14s 5.74% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func2
0 0% 98.36% 0.27s 11.07% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func4
0 0% 98.36% 1.15s 47.13% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].unbuffer[go.shape.int].func6
0 0% 98.36% 1.15s 47.13% github.com/achille-roussel/kway-go.benchmark[go.shape.int]
0 0% 98.36% 0.76s 31.15% iter.Pull2[go.shape.[]go.shape.int,go.shape.interface { Error string }].func1
0 0% 98.36% 0.76s 31.15% runtime.corostart
As we can see here, a significant amount of time seems to be spent in the Go runtime code managing coroutines. While it might be possible to optimize the runtime, there is a lower bound on how much it can be reduced.
It is also unlikely that the Go compiler could help here, there are no real opportunities for inlining or other optimizations.
We basically have a very high baseline cost for each operation, with the hypothesis that it is driven by coroutine context switch implemented in the runtime, the only thing we can do to improve performance is doing less of these.
This is a typical a baseline cost amortization problem: we want to call the
next
function returned by iter.Pull2
less often, which can be done by
introducing buffering. Instead of pulling values one at a time, we can
efficiently buffer N values from each sequence in memory, by transposing
the iter.Seq2[T, error]
sequences into iter.Seq2[[]T, error]
. The call
to next
then only needs to happen when we exhaust the buffer, which ends up
amortizing its cost.
With an internal buffer size of 128 values per sequence:
Merge2 190103247 6.133 ns/op 0.8333 comp/op 163045156 merge/s
Merge3 95485022 12.74 ns/op 1.864 comp/op 78492807 merge/s
Now we made the algorithm 3-4x faster, and have performance in the range of 1.5 to 2.5x the theoretical throughput limit.
It is interesting to note that the CPU profile didn't seem to indicate that 75% of the time was spent in the runtime, but reducing the time spent in that code path has had a non-linear impact on performance. Likely some other CPU instruction pipeline and caching shenanigans are at play here, possibly impacted by the atomic compare-and-swap operations in coroutine switches.
As expected, the CPU profile now shows that almost no time is spent in the runtime:
Duration: 3.17s, Total samples = 2.35s (74.08%)
Showing nodes accounting for 2.28s, 97.02% of 2.35s total
Dropped 22 nodes (cum <= 0.01s)
flat flat% sum% cum cum%
0.45s 19.15% 19.15% 0.56s 23.83% github.com/achille-roussel/kway-go.(*tree[go.shape.int]).next
0.43s 18.30% 37.45% 0.43s 18.30% github.com/achille-roussel/kway-go.benchmark[go.shape.int].func2
0.37s 15.74% 53.19% 0.97s 41.28% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].merge2[go.shape.int].func3
0.23s 9.79% 62.98% 0.24s 10.21% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func1.1
0.22s 9.36% 72.34% 0.65s 27.66% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].unbuffer[go.shape.int].func6.1
0.13s 5.53% 77.87% 0.13s 5.53% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func4.1
0.12s 5.11% 82.98% 0.21s 8.94% github.com/achille-roussel/kway-go.benchmark[go.shape.int].func1
0.10s 4.26% 87.23% 0.52s 22.13% github.com/achille-roussel/kway-go.sequence.func1
0.09s 3.83% 91.06% 0.09s 3.83% cmp.Compare[go.shape.int] (inline)
0.05s 2.13% 93.19% 0.05s 2.13% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func2.1
0.03s 1.28% 94.47% 0.06s 2.55% runtime/pprof.(*profMap).lookup
0.02s 0.85% 95.32% 0.02s 0.85% github.com/achille-roussel/kway-go.parent (inline)
0.02s 0.85% 96.17% 0.02s 0.85% runtime.asyncPreempt
0.02s 0.85% 97.02% 0.02s 0.85% runtime.mapaccess1_fast64
0 0% 97.02% 0.97s 41.28% github.com/achille-roussel/kway-go.BenchmarkMerge2
0 0% 97.02% 0.76s 32.34% github.com/achille-roussel/kway-go.BenchmarkMerge3
0 0% 97.02% 0.31s 13.19% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func1
0 0% 97.02% 0.08s 3.40% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func2
0 0% 97.02% 0.13s 5.53% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].buffer[go.shape.int].func4
0 0% 97.02% 0.76s 32.34% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].merge[go.shape.int].func5
0 0% 97.02% 1.73s 73.62% github.com/achille-roussel/kway-go.MergeFunc[go.shape.int].unbuffer[go.shape.int].func6
0 0% 97.02% 1.73s 73.62% github.com/achille-roussel/kway-go.benchmark[go.shape.int]
0 0% 97.02% 0.52s 22.13% iter.Pull2[go.shape.[]go.shape.int,go.shape.interface { Error string }].func1
0 0% 97.02% 0.52s 22.13% runtime.corostart
There is a final performance frontier we can cross. While we are buffering
values internally, the input and output sequences remain iter.Seq2[T, error]
,
which yield values one by one. Often times in data systems, APIs have pagination
capabilities, or stream processors work on batch of values for the same reason
we added buffering: it reduces the baseline cost of crossing system boundaries.
If the input sequences are already slices of values, and the output sequence produces slices of values, we can reduce the internal memory footprint (no need to allocate memory to buffer the inputs), while also further amortizing the cost of function calls to yield values in and out of the merge algorithm.
Applications that fall into those categories can unlock further performance by
using MergeSlice
instead of Merge
, which works on iter.Seq2[[]T, error]
end-to-end.
What is interesting with this approach is that in cases where the processing of inputs and outputs can be batched, this model can even beat the theoretical throughput limit. For example, in the benchmarks we've used, the body of the loop consuming merged values simply counts the results. When consuming slices there is no need to iterate over the slices and increment the counter by one each time, we can batch the operation by incrementing the counter by the length of the slice, achieving much higher throughput than predicted by the baseline:
MergeSlice2 477720793 2.273 ns/op 0.6688 comp/op 439971259 merge/s
MergeSlice3 150406080 7.945 ns/op 1.667 comp/op 125861613 merge/s
⚠️ Keep in mind that to minimize the footprint,MergeSlice
resuses its output buffer, which means that the application cannot retain it beyond the body of the loop raning over the merge function. This can lead to subtle bugs that can be difficult to track,Merge
should always be preferred unless there is clear evidence that the increased maintenance cost is worth the performance benefits.