feat: Minimal op registration in isolate

[bartlomieju]

Towards #2987

This PR introduces OpRegistry structure in Isolate.

OpRegistry collects OpHandlers which are functions returning CoreOp.

This can be safely introduced to work with deno_cli because if you set dispatch manually via set_dispatch then ops registered on isolate are ignored.

I still have to get op id forwarding to JS.

CC @ry @piscisaureus

[bartlomieju]

$ cargo run --release --example deno_core_http_bench

third_party/wrk/mac/wrk -d 10s --latency http://127.0.0.1:4544
Running 10s test @ http://127.0.0.1:4544
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   121.98us   41.94us   2.00ms   93.58%
    Req/Sec    39.43k     2.08k   41.72k    82.18%
  Latency Distribution
     50%  114.00us
     75%  119.00us
     90%  142.00us
     99%  292.00us
  792454 requests in 10.10s, 38.54MB read
Requests/sec:  78458.19
Transfer/sec:      3.82MB

$  cargo run --release --example deno_core_http_bench -- --multi-thread

third_party/wrk/mac/wrk -d 10s --latency http://127.0.0.1:4544
Running 10s test @ http://127.0.0.1:4544
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   185.81us  613.64us  18.64ms   98.18%
    Req/Sec    36.48k     6.32k   43.68k    87.13%
  Latency Distribution
     50%  106.00us
     75%  142.00us
     90%  228.00us
     99%    1.22ms
  733098 requests in 10.10s, 35.66MB read
Requests/sec:  72578.50
Transfer/sec:      3.53MB

[kevinkassimo]

$ cargo run --release --example deno_core_http_bench

third_party/wrk/mac/wrk -d 10s --latency http://127.0.0.1:4544
Running 10s test @ http://127.0.0.1:4544
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   121.98us   41.94us   2.00ms   93.58%
    Req/Sec    39.43k     2.08k   41.72k    82.18%
  Latency Distribution
     50%  114.00us
     75%  119.00us
     90%  142.00us
     99%  292.00us
  792454 requests in 10.10s, 38.54MB read
Requests/sec:  78458.19
Transfer/sec:      3.82MB

$  cargo run --release --example deno_core_http_bench -- --multi-thread

third_party/wrk/mac/wrk -d 10s --latency http://127.0.0.1:4544
Running 10s test @ http://127.0.0.1:4544
  2 threads and 10 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   185.81us  613.64us  18.64ms   98.18%
    Req/Sec    36.48k     6.32k   43.68k    87.13%
  Latency Distribution
     50%  106.00us
     75%  142.00us
     90%  228.00us
     99%    1.22ms
  733098 requests in 10.10s, 35.66MB read
Requests/sec:  72578.50
Transfer/sec:      3.53MB

Do you have the comparison with the master branch?

[bartlomieju]

@kevinkassimo

master

"req_per_sec": {
    "deno_tcp": 53518, 
    "deno_core_single": 54066, 
    "node_proxy_tcp": 20393, 
    "deno_proxy_tcp": 18601, 
    "deno_core_multi": 58185, 
    "deno_tcp_current_thread": 51059, 
    "hyper": 63312, 
    "node_tcp": 54550, 
    "deno_http": 18684, 
    "node_http": 19992, 
    "deno_proxy": 990, 
    "node_proxy": 2640
  }, 
  "max_latency": {
    "deno_tcp": 9.04, 
    "deno_core_single": 0.36, 
    "node_proxy_tcp": 1.05, 
    "deno_proxy_tcp": 5.36, 
    "deno_core_multi": 0.95, 
    "deno_tcp_current_thread": 9.46, 
    "hyper": 2.86, 
    "node_tcp": 0.597, 
    "deno_http": 2.62, 
    "node_http": 2.22, 
    "deno_proxy": 19.51, 
    "node_proxy": 9.4
  }, 

this PR

"req_per_sec": {
    "deno_tcp": 51090, 
    "deno_core_single": 58520, 
    "node_proxy_tcp": 19952, 
    "deno_proxy_tcp": 18476, 
    "deno_core_multi": 61798, 
    "deno_tcp_current_thread": 48471, 
    "hyper": 63395, 
    "node_tcp": 53003, 
    "deno_http": 18526, 
    "node_http": 19925, 
    "deno_proxy": 961, 
    "node_proxy": 2533
  }, 
  "max_latency": {
    "deno_tcp": 8.89, 
    "deno_core_single": 0.405, 
    "node_proxy_tcp": 1.06, 
    "deno_proxy_tcp": 5.62, 
    "deno_core_multi": 1.07, 
    "deno_tcp_current_thread": 9.13, 
    "hyper": 2.56, 
    "node_tcp": 0.588, 
    "deno_http": 2.72, 
    "node_http": 2.28, 
    "deno_proxy": 18.11, 
    "node_proxy": 10.16
  }, 

That's a nice speed up in req/secs, but latency is slightly worse.

[ry]

Looks good! This is what I had in mind :)

[ry]

Bartek - proxies shouldn’t be involved - they may kick us out of optimizations. I guess that Deno.ops[“foo”] should be an explicit function call instead. Maybe Deno.ops.get(“foo”)?

[bartlomieju]

Bartek - proxies shouldn’t be involved - they may kick us out of optimizations. I guess that Deno.ops[“foo”] should be an explicit function call instead. Maybe Deno.ops.get(“foo”)?

Aaaah I was worried that's not the best solution. Yes, Deno.ops.get("foo") is the way to go, updating.

EDIT: Although I'm not quite convinced it should be Deno.ops instead of Deno.core.ops. Why introduce another global namespace?

[ry]

LGTM
Nice work! Let’s watch to see how this effects the benchmarks of deno_core_single and deno_core_multi