usm: map cleaner: Use batch operations

[guyarb]

What does this PR do?

Migrates map cleaner to use batch operations.
The new method will allow us to reduce the cost of the map cleaner (performance wise) and reduce the number of syscalls we perform.

Motivation

Using batch operations allowing us to cut allocations, memory pressure and general runtime by 50%.

Additional Notes

Benchmarks:

main:
BenchmarkCleaner1000
BenchmarkCleaner1000-16      	     342	   3606537 ns/op	  271356 B/op	   10127 allocs/op
BenchmarkCleaner10000
BenchmarkCleaner10000-16     	      39	  28976853 ns/op	 2964974 B/op	  104637 allocs/op
BenchmarkCleaner100000
BenchmarkCleaner100000-16    	       3	 408697443 ns/op	32297042 B/op	 1049653 allocs/op

new branch:
With batches:
BenchmarkBatchCleaner1000Entries10PerBatch
BenchmarkBatchCleaner1000Entries10PerBatch-16        	     610	   2111341 ns/op	  148518 B/op	    5737 allocs/op
BenchmarkBatchCleaner1000Entries100PerBatch
BenchmarkBatchCleaner1000Entries100PerBatch-16       	     654	   1885137 ns/op	  147005 B/op	    5192 allocs/op
BenchmarkBatchCleaner10000Entries100PerBatch
BenchmarkBatchCleaner10000Entries100PerBatch-16      	      66	  18089427 ns/op	 1450936 B/op	   55235 allocs/op
BenchmarkBatchCleaner10000Entries1000PerBatch
BenchmarkBatchCleaner10000Entries1000PerBatch-16     	      63	  17859127 ns/op	 1459255 B/op	   54696 allocs/op
BenchmarkBatchCleaner100000Entries100PerBatch
BenchmarkBatchCleaner100000Entries100PerBatch-16     	       6	 193079698 ns/op	14796210 B/op	  555677 allocs/op
BenchmarkBatchCleaner100000Entries1000PerBatch
BenchmarkBatchCleaner100000Entries1000PerBatch-16    	       6	 207960644 ns/op	14642945 B/op	  550267 allocs/op

without batches:
BenchmarkCleaner1000Entries
BenchmarkCleaner1000Entries-16                       	     331	   3053879 ns/op	  180851 B/op	    8126 allocs/op
BenchmarkCleaner10000Entries
BenchmarkCleaner10000Entries-16                      	      37	  30940045 ns/op	 1823625 B/op	   84630 allocs/op
BenchmarkCleaner100000Entries
BenchmarkCleaner100000Entries-16                     	       4	 300761211 ns/op	18694620 B/op	  849658 allocs/op

Batches - improve memory, allocations and runtime by 50%.
Without batches - Improves allocations by 20%, memroy by 33%, and runtime by 15%.

Possible Drawbacks / Trade-offs

Describe how to test/QA your changes

Reviewer's Checklist

• If known, an appropriate milestone has been selected; otherwise the Triage milestone is set.
• Use the major_change label if your change either has a major impact on the code base, is impacting multiple teams or is changing important well-established internals of the Agent. This label will be use during QA to make sure each team pay extra attention to the changed behavior. For any customer facing change use a releasenote.
• A release note has been added or the changelog/no-changelog label has been applied.
• Changed code has automated tests for its functionality.
• Adequate QA/testing plan information is provided if the qa/skip-qa label is not applied.
• At least one team/.. label has been applied, indicating the team(s) that should QA this change.
• If applicable, docs team has been notified or an issue has been opened on the documentation repo.
• If applicable, the need-change/operator and need-change/helm labels have been applied.
• If applicable, the k8s/<min-version> label, indicating the lowest Kubernetes version compatible with this feature.
• If applicable, the config template has been updated.

[pr-commenter]

Bloop Bleep... Dogbot Here

Regression Detector Results

Run ID: a6707cbb-2a67-4215-943f-0452d734d53d
Baseline: b7403de
Comparison: 1e99e6a
Total datadog-agent CPUs: 7

Explanation

A regression test is an integrated performance test for datadog-agent in a repeatable rig, with varying configuration for datadog-agent. What follows is a statistical summary of a brief datadog-agent run for each configuration across SHAs given above. The goal of these tests are to determine quickly if datadog-agent performance is changed and to what degree by a pull request.

Because a target's optimization goal performance in each experiment will vary somewhat each time it is run, we can only estimate mean differences in optimization goal relative to the baseline target. We express these differences as a percentage change relative to the baseline target, denoted "Δ mean %". These estimates are made to a precision that balances accuracy and cost control. We represent this precision as a 90.00% confidence interval denoted "Δ mean % CI": there is a 90.00% chance that the true value of "Δ mean %" is in that interval.

We decide whether a change in performance is a "regression" -- a change worth investigating further -- if both of the following two criteria are true:

1. The estimated |Δ mean %| ≥ 5.00%. This criterion intends to answer the question "Does the estimated change in mean optimization goal performance have a meaningful impact on your customers?". We assume that when |Δ mean %| < 5.00%, the impact on your customers is not meaningful. We also assume that a performance change in optimization goal is worth investigating whether it is an increase or decrease, so long as the magnitude of the change is sufficiently large.

2. Zero is not in the 90.00% confidence interval "Δ mean % CI" about "Δ mean %". This statement is equivalent to saying that there is at least a 90.00% chance that the mean difference in optimization goal is not zero. This criterion intends to answer the question, "Is there a statistically significant difference in mean optimization goal performance?". It also means there is no more than a 10.00% chance this criterion reports a statistically significant difference when the true difference in mean optimization goal is zero -- a "false positive". We assume you are willing to accept a 10.00% chance of inaccurately detecting a change in performance when no true difference exists.

The table below, if present, lists those experiments that have experienced a statistically significant change in mean optimization goal performance between baseline and comparison SHAs with 90.00% confidence OR have been detected as newly erratic. Negative values of "Δ mean %" mean that baseline is faster, whereas positive values of "Δ mean %" mean that comparison is faster. Results that do not exhibit more than a ±5.00% change in their mean optimization goal are discarded. An experiment is erratic if its coefficient of variation is greater than 0.1. The abbreviated table will be omitted if no interesting change is observed.

No interesting changes in experiment optimization goals with confidence ≥ 90.00% and |Δ mean %| ≥ 5.00%.

Fine details of change detection per experiment.

experiment	goal	Δ mean %	Δ mean % CI	confidence
tcp_syslog_to_blackhole	ingress throughput	+0.68	[+0.55, +0.81]	100.00%
file_tree	egress throughput	+0.23	[-1.61, +2.07]	16.25%
dogstatsd_string_interner_8MiB_50k	ingress throughput	+0.02	[-0.01, +0.06]	74.34%
dogstatsd_string_interner_8MiB_100k	ingress throughput	+0.02	[-0.02, +0.07]	63.44%
trace_agent_json	ingress throughput	+0.02	[-0.12, +0.15]	16.05%
uds_dogstatsd_to_api	ingress throughput	+0.01	[-0.16, +0.19]	10.74%
dogstatsd_string_interner_8MiB_100	ingress throughput	+0.00	[-0.12, +0.13]	3.59%
dogstatsd_string_interner_128MiB_1k	ingress throughput	+0.00	[-0.14, +0.14]	0.39%
dogstatsd_string_interner_64MiB_1k	ingress throughput	-0.00	[-0.13, +0.13]	0.18%
dogstatsd_string_interner_64MiB_100	ingress throughput	-0.00	[-0.14, +0.14]	0.27%
dogstatsd_string_interner_128MiB_100	ingress throughput	-0.00	[-0.14, +0.14]	0.45%
idle	egress throughput	-0.00	[-2.46, +2.46]	0.03%
trace_agent_msgpack	ingress throughput	-0.00	[-0.13, +0.13]	0.65%
dogstatsd_string_interner_8MiB_10k	ingress throughput	-0.00	[-0.01, +0.01]	21.70%
file_to_blackhole	egress throughput	-0.00	[-1.02, +1.01]	0.59%
dogstatsd_string_interner_8MiB_1k	ingress throughput	-0.01	[-0.11, +0.09]	13.37%
tcp_dd_logs_filter_exclude	ingress throughput	-0.03	[-0.17, +0.11]	26.55%
otel_to_otel_logs	ingress throughput	-0.24	[-1.83, +1.34]	19.95%

[brycekahle]

Where are the allocation savings coming from? You still have to accumulate all the keys you want to delete.

[p-lambert]

Please re-add the comment (not necessarily here) as this is still relevant. You're still passing a unsafe.Pointer to MapIterator.Next().
The reason why this was done originally was to avoid the marshalling cost as noted in the comment.

[p-lambert]

it's really nice to see a bunch of hacks go away now that we have generics :)

[guyarb]

Also, the use of generics introduced improvement, as we don't need marshalBytes!

[p-lambert]

Do we really want to execute this in a loop? I think it's not that unlikely that in a real-world workload this could run forever which would cause the MapCleaner to hang during termination, no?
In this pathological scenario you would supply also a stale timestamp to the callback

[guyarb]

Not sure about running forever / adding timed context here
but instead, we can "halt" if we processed mc.emap.MaxEntries
what do you think about that?

Along side with aborting the loop when we got 0 entries from the batch api, it will guarantee we won't be stuck forever

[p-lambert]

sounds reasonable to me 👍

[p-lambert]

nit/suggestion: let's try to be more strict with the type constraints here. For example, we don't want pointers, interfaces etc since this would cause issues with the code downstream.

[guyarb]

Tried to look for a way of doing that, but couldn't find such a way (to exclude interface / pointer)
if you have a way - I'd be happy to learn
having said that - we still have a guard - UTs

[guyarb]

Where are the allocation savings coming from? You still have to accumulate all the keys you want to delete.

In the original impl (a.k.a iterator) ~90% of the allocations are coming from MapIterator.Next

In the batches api - there are much less allocations from cillium

And the improvement we have between the without batching algorithm compared to main, is achieved by using generics, which made the direct call of marshalBytes redundant