Refactor exporter - step 4

[reyang]

This is a follow up PR of #1078.

Changes

• Added a multi-writer single-reader circular buffer, which will be used for batch exporting scenario. Key points:
  • The entire buffer is pre-allocated.
  • Consumer doesn't need to copy items to a new array/list, the consumption is happening "in-place".
• Still need to read more about the CLR/JIT memory model (about the Int64 operation) to see if additional memory barrier should be added.
• I'll do some initial perf test offline, additional optimization can be done later without having the affect the interface.
• IEnumerator optimization will be covered separately in Perf improvement using zero-alloc foreach #1080.
• [Don't merge] Stress test for CircularBuffer #1086 has the stress test code FYI.

For significant contributions please make sure you have completed the following items:

• Design discussion issue #
• Changes in public API reviewed

[codecov]

Codecov Report

Merging #1085 into master will decrease coverage by 0.42%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##           master    #1085      +/-   ##
==========================================
- Coverage   76.59%   76.17%   -0.43%     
==========================================
  Files         224      225       +1     
  Lines        6165     6199      +34     
==========================================
  Hits         4722     4722              
- Misses       1443     1477      +34     

Impacted Files	Coverage Δ
src/OpenTelemetry/Internal/CircularBuffer.cs	0.00% <0.00%> (ø)

[CodeBlanch]

I know it's a really really big number, but in theory it could wrap around to negative at some point which would throw overflow exception I think? If we switched to ulong head/tail and did unchecked (this.head = headSnapshot + 1) I think the wrap would be fine given the % this.capacity in the algorithm?

[reyang]

I guess it is impractical to hit that number? Adding the wrapping support requires more than unchecked, it will touch other part of the lock-free algorithm and put additional requirements on memory barriers. So I would prefer to keep it as-is in favor of perf vs. mathematical correctness.

[timr-dev]

Do you want to provide an additional check with a boolean and an Interlock to prevent 2 threads from accidently entering if the single reader pattern is accidently broken.

[ThomsonTan]

Is headSnapshot validation needed, or Interloaded.Add can work here?

[reyang]

It's is required, otherwise it'll break the 🧙‍♀️.

[ThomsonTan]

I see, yes, it is required because the item pointed by headSnapshot maybe not in the range of [this.tail, this.head).

[CodeBlanch]

LGTM

[ThomsonTan]

As the name implied it is a circular buffer, should it TryAdd overwrite the oldest item when the buffer is full instead of returning false on the most recent item?

[ThomsonTan]

Retry here introduces uncertainty of running time of TryAdd, perhaps return false instead of retry in the while(true) loop, or limit retry count?

[reyang]

The intention of this function is to return false only when queue is full.

[reyang]

@ThomsonTan you've actually worried me! 😅 Here goes a PR #1088 to address it.

[ThomsonTan]

Restrict capacity to be power of 2, so the modulo operation % could replaced with a simple bitwise and?

[reyang]

How much (X%) perf improvement would you expect from modulo?

[ThomsonTan]

div could take tens times of CPU cycles than bitwise arithmetic instruction, but there is also trade off taking more memory (for ceiling(2**size)). This could be a low priority optimization and needs data to prove.

[ThomsonTan]

When the queue is empty, wait for notification of new items instead of polling?

[reyang]

That would make it a normal lock based queue instead of lock-free.