Better ReferenceCounter branch[tarjin-2]

[erikzhang]

(Maybe) an optimized version of #459

[erikzhang]

We need to test which one is better, this or #459?

[dusmart]

test with old PoC VgEMAkABSotKi0qLSotKi0qLSotKi0qLSotKiwH4D40MAkAAixHAAQAEYFidYEofDAtkZXNlcmlhbGl6ZQwUwO85zuDk6SXGwqBqeeFEDdhvzqxBYn1bUkVYJNFJ from Tanya-6

tarjin-2 cost 0:00:07
tarjan cost 0:00:09
no-optimization cost 0:05:26
tarjin-3-switch 0:00:04
optimize-reference-counter 0:00:06

it's a huge improvement

[dusmart]

test with a PoC derived from Tanya-4 whose itration number is 3800 instead of 380000 (otherwise the time will be longer)

VwEAwkpKAfsHdwARwG8AnXcAbwAl9////xHAzwHYDncAEM5vAJ13AG8AJff///9F

tarjan => 0:00:21
tarjan-2 => 0:00:15
no-optimization => 0:00:01

Tanya-4 and Tanya-6 is the only two issues which have not been fixed thoroughly.

if we set iteration number to 38000 which is VwEAwkpKAfsHdwARwG8AnXcAbwAl9////xHAzwJwlAAAdwAQzm8AnXcAbwAl9////0U=

tarjan => 0:01:09
tarjan-2 => 0:02:01
no-optimization => 0:00:12
tarjan-3-swtich => 0:00:27
optimize-reference-counter => 0:00:49

I'm trying my best to do the time complexity analysis ...... This result is weird for me

[dusmart]

Here are the differences between Tanya-4 and Tanya-6.

Poc	Tanya-4	Tanya-6
iteration op	PICKITEM	deserialize+drop
iteration num	380000	1024
operation target	[[[ ...2044... ]]] end-to-end loop array	[[[ ...2046... ]]] simple nested array
best GC	no-optimization	tarjin-2
no-optimization time complexity	O(n)	O(n2)
no-optimization graph
tarjan-2 time complexity	O(n)	O(n)
tarjan-2 graph	tarjan is O(E+V)	the same as left

[dusmart]

Actually, to prevent Tanya-6, we only need to optimize the AddReference like this:

I've tested this with Tanya-6, 0:00:02 cost.

[dusmart]

The best time complexity for GC is O(n), therefore Tanya-4 is hard to prevent. And at this time, there are lots of other O(n) issues such as Bytes opration. Just ignore Tanya-4 will be an understandable choice.

In a certain point of view, neo-project/neo#2725 is also a variant of Tanya-4, the GetNotification syscall creates lots of small object and the dropping will cause a slow GC.

[erikzhang]

@dusmart I've made some improvements based on tarjin-2. Please check #461

[dusmart]

@dusmart I've made some improvements based on tarjin-2. Please check #461

Tanya-6 cost 0:00:06 which is still a huge improvement compared to no-optimization.
Tanya-4 with 38000 iteration cost 0:01:25 which is similar to tarjin-2.

The result is based on the cheapest MacBookAir with m1 chip and release version of dotnet6.0.
The result is not that stable.

After testing many times with Tanya-4 and iteration number set to 10000, tarjin-3 cost 0:00:25 while tarjin-2 cost 0:00:34 adn tarjan cost 0:00:18. They all have a O(n) time complexity in this case.

[erikzhang]

#463 is the best. Closed this one.