Performance of PER codec

[dudycz]

Hi. I have been playing with two most popular PER codecs: rasn and asn1-codecs and made some benchmark comparing their performance. One thing I have noticed is that rasn can be ~10x slower in encoding in some complex cases:

Codec	Encoding (µs)	Decoding (µs)
rasn	1916.7	826.8
asn1-codecs	208	145.5

I looked into flamegraphs and callgrinds but I couldn't figure out what contributes to this big difference. If you're interested I could try to collect some and attach here.

Link to repo with benchmark: https://github.com/dudycz/asn1_codecs_bench

[XAMPPRocky]

Thank you for your issue! Would you be able to attach the flamegraphs? That would be very helpful. In general the initial version I made prioritised correctness over performance so there's likely a lot of places where it could be more efficient. I already know that I haven't added the empty struct optimisation that was added to BER to PER.

[dudycz]

I had to encode sample.asn 256 times to get something recorded in flamegraph. Here it is:

[Nicceboy]

Seems like your test bench uses mostly integers. asn1-codecs crate seems to use i128 internally for Integer type while rasn uses BigInts. I would guess that this plays big factor at least.

I think there was a plan to make inner implementation of integer types as feature, so that it would be possible to select different internal type for performance reasons if very big numbers are not required.

[XAMPPRocky]

Yeah, that's still a todo, I should write up an issue giving details in case someone else who has more time is interested.

[Nicceboy]

Yeah, that's still a todo, I should write up an issue giving details in case someone else who has more time is interested.

I think this is rather important optimisation in general and should not take too much work to implement, if we have, for example, just the i128 and BigInts as initial options. If you have time to write what you had on your mind, maybe I can try to do it.

[Nicceboy]

I have been reworking integer type (by using primitives (i128) by default, and switching to larger ones on overflows, or if big one is created manually) Hopefully I can open draft PR in the end of week. The resulting integer will be enum, and the type of the big integers does not matter that much anymore.

Not sure if it is the best approach, but it is the one I chose after trying quite many different things. Let's leave those comments for the PR. It can be completely changed still.

Seems like integers are not the only problem with UPER. The extensive use of new vector buffers and moving this data contributes more. Default low capacity's in vectors, a lot single pushes and overall creation of new buffers instead of sharing pointer of one or reusing existing allocations slows down quite much.

Some initial differences on M2 Pro from the integer change:

UPER:

COER (The difference was much more impactful)

I have also made initial rework for optimising allocations in COER (the results below are based on the int remake) Maybe UPER will follow if I have time.

So by changing the integer type and reducing allocations, it was already possible to get 3x speedup at least for COER, based on the benchmark base of @dudycz .

Allocations could be improved further but I am having painful issues with lifetimes.

[repnop]

@dudycz since #256 got merged a little while ago, are you able to rerun your benchmark, at least for the decoding side of things? it should be quite a bit better now.

[dudycz]

Yes, I can confirm big improvement in decoding time (on my "bench" setup I observed decrease from 457us to 341us!). Good job!

[dudycz]

Meanwhile I had updated my benchmark with third uper codec - asn1rs.

Codec	Encoding (µs)	Decoding (µs)
rasn	1759	259
asn1-codecs	187	122
asn1rs	72	65

[XAMPPRocky]

Separately I think we should add continuous profiling to the CI and I've created an issue for that. Since this issue doesn't have a specific goal/end, I'm going to move this to a discussion.

#302