[For Testing] Bundled Memory Changes

[benaadams]

Don't alloc header names #391
Consume body without allocating or reading #335 (PR has changed)
Reduce GetString allocs and conversions #312 (PR has changed)
ConcurrentStack -> ConcurrentQueue #390
GC.Collect before pinning starts #331
Dispose of MemoryPool2; suppress its finalizers #343
Pre-allocate memory pool #362
Initialize the work queue sizes #363
Precomputed header bytes #367
OnCompleted doesn't need threadpool #380

[benaadams]

Also resolved some slight conflicts

[benaadams]

Now the difference in allocations between ConcurrentQueue and ConcurrentStack are showing; swapping MemoryPool

[benaadams]

Pre Allocs

Post Allocs

Not sure about stack vs queue; reverting and will hold on to block for longer

[benaadams]

Gone back to Queue rather than Stack

Stack allocs

Queue allocs

[NickCraver]

I setup benchmarks on the Stack Overflow web tier to give us a good idea of what our actual performance will be and @benaadams was kind enough to bundle the changes. Here's current aspnet/dev:

-bash-4.2$ ./wrk -s pipeline.lua -t32 -c256 -d10s http://co-web11:5001/plaintext -- 16
Running 10s test @ http://co-web11:5001/plaintext
  32 threads and 256 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    20.98ms   61.39ms 906.56ms   92.04%
    Req/Sec    22.08k     4.13k   73.21k    88.44%
  6969114 requests in 10.10s, 0.86GB read
Requests/sec: 690161.21
Transfer/sec:     86.88MB

...and here's this branch:

-bash-4.2$ ./wrk -s pipeline.lua -t32 -c256 -d10s http://co-web11:5001/plaintext -- 16
Running 10s test @ http://co-web11:5001/plaintext
  32 threads and 256 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    29.87ms   71.90ms 479.58ms   83.73%
    Req/Sec    38.22k     8.93k   79.89k    79.82%
  12034366 requests in 10.10s, 1.48GB read
Requests/sec: 1191674.89
Transfer/sec:    150.01MB

[benaadams]

Almost there, down another 719,992 bytes and 8,000 objects

[NickCraver]

Updating numbers, regression on the memory pool commit (look at latency specifically) and we're hitting 100% CPU on the web server now (didn't get comparison CPU before the rebase/force):

Running 10s test @ http://co-web11:5001/plaintext
  32 threads and 256 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    62.12ms  100.42ms 946.49ms   78.24%
    Req/Sec    35.08k    11.44k   86.41k    73.63%
  10963259 requests in 10.10s, 1.35GB read
Requests/sec: 1085653.38
Transfer/sec:    136.67MB

[benaadams]

I removed a commit is the middle of the chain and rebasing got upset; all the other commits are individually open as PRs and referenced at the top; so will ignore the last two for now :)

[benaadams]

Stashed the "experi-mental" changes; and just gone for the merges - hopefully back to previous pref...

[benaadams]

Other changes (SocketOutput taking MemoryPool2 inputs) here https://github.com/benaadams/KestrelHttpServer/tree/so-testing-exotic

[AspNetSmurfLab]

Benchmark results from the SmurfLab:

Running 10s test @ http://10.0.0.100:5001/plaintext
  32 threads and 256 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     6.56ms   14.11ms 264.13ms   94.59%
    Req/Sec    20.66k     2.35k   28.12k    81.86%
  6496342 requests in 10.10s, 817.79MB read
  Socket errors: connect 0, read 0, write 600, timeout 0
Requests/sec: 643225.30
Transfer/sec:     80.97MB
``

[benaadams]

Closing this, as testing done