[mORMot] speed up pipelining; speed up docker build #8057

pavelmash · 2023-03-21T15:04:08Z

mORMot using a modified libpq that allows to send the PipelineSync command without flushing prev. commands to socket

…command without flushing prev. commands to socket

sebastienros · 2023-03-22T00:11:47Z

frameworks/Pascal/mormot/src/raw.pas

@@ -193,6 +193,7 @@ constructor TRawAsyncServer.Create(
     hsoIncludeDateHeader  // required by TPW General Test Requirements #5
    ] + flags);
  fHttpServer.HttpQueueLength := 10000; // needed e.g. from wrk/ab benchmarks
+  fHttpServer.ServerName := 'M'; // better plaintext performance


This optimization has been disallowed for other frameworks by the past:

We expect the Server header to be whatever is normal for the platform or framework. If the framework does not normally provide a Server response header, we nevertheless require that you provide one as this roughly normalizes network load across all implementations.

Default Server: header produced by mORMot2 framework is mORMot2 (Linux) what is too long. So we override it to short one-letter version, as most other top-rated frameworks did, for example - .net use "K" instead of "Kestrel"

@pavelmash
Your server already has very very good performance: 6959479 req/sec.
Не занимайтесь ерундой.

Must have been confused with the /p in plaintext routes. Thanks for confirming.

It's disallowed to all play with the same rules.
But still not formalized in the rules and verification tests. :(

Yes, I know.
It's for that, than we need to clarify it and add rule and verification.

For example, with Nginx you can't change the server name.

Server header value is changed to mmt by [658165b] .
@joanhey - hope rule will be changed - I propose to use at last 3 character. For example google respond with server: gws, so 4 characters is too much.
And, for sure, we should add validation for new rule, to ensure all participants are in the same situation

We need to clarify this, today socketify.py uses Server: uWebSockets_20, we can remove Server header or add a minimal size like 3 characters, in my case I can change to Server: uWS

@joanhey

For example, with Nginx you can't change the server name.

so you need to add additional characters to the response body to be honest

sebastienros · 2023-03-22T00:13:40Z

frameworks/Pascal/mormot/src/raw.pas

@@ -233,6 +234,7 @@ function TRawAsyncServer.GetRawRandomWorlds(cnt: PtrInt; out res: TWorlds): bool
    pStmt.SendPipelinePrepared;
    pConn.PipelineSync;
  end;
+  pConn.Flush; // we use modified libpq what not flush inside PQPipelineSync - flush manually


Sounds like a discussion that concluded by the requirement of SYNC calls. I would suggest you double check before using a modified libpq that prevents them.

I checked. Requirements is to send a SYNC command, but not to flush client socket (do write syscall) after SYNC.
So we modify a lipbq to avoid socket flushing - see what we did.
At last Just-JS do the same (send all pipeline commands in one-two socket write) and it's implementation is valid

@sebastienros - checked once more using Postgres protocol verification tool as suggested in this post

Everything is green.

@pavelmash great then, thanks for taking the time to double-check and sorry for the inconvenience. I had forgotten about the CI verification step.

IMO will be forbidden to use a modified libpq.
We are testing frameworks !!

FYI my libpq patch is available here.

FYI

anyone should use any connectors, after all, optimizations are optimizations. The bad thing is to try to compare WebFrameworks with databases instead of direct testing the database connectors/solutions.

I don't think DB tests should be separated. Optimizations are indeed optimizations.
The real bad thing is that we compare frameworks which are proofs of concept, not usable in real-purpose code, with production ready frameworks.
For instance, some frameworks use a direct DB connection, but only handle integer fields, so that they could shine in TFB benchmarks. Or others don't allow to change the fortunes template without re-compiling the executable - which IMHO voids the aim of a HTML template, which is to separate view and model.
I would add a requirement with some complex process instead, something with data more complex than the /fortunes test, just to validate the framework is usable for any real usecase, i.e. is a true "realistic" framework.

It's better to create a general issue or discussion, than talk in a PR.
We need to try to have a fair play, for all frameworks.
And later create individual issues, for each specific item.

@synopse not only the HTML template need compilation. I see some frameworks than cache the response in Fortunes, so bypassing the template, composition and html escape stage in each request.
We need optimizations, but general, not specific only to the benchmark.

PD: @pavelmash you can use only 1 letter server header, till is NOT added as a rule and verification.

@joanhey - reverted by [1a48475] to use 1 letter, thanks!
I agree what separate issue should be created for discussion.

Hope this PR will be merged now. At last this can help Postgres team to estimate changes we propose to libpq what implemented by @volyrique as libpq patch. In case Postgres team decline such patch I will rollback usage of modified libpq

…ebate

…verification

* [mORMot] using a modified libpq that allows to send the PipelineSync command without flushing prev. commands to socket * [mORMot] fix dockerfile, speed up a docker build command * [mORMot] shorten server name for better plaintext performance * [mORMot] change `Server` header value to `mmt` because `M` provokes debate * [mORMot] use 1 letter server header, till is NOT added as a rule and verification --------- Co-authored-by: pavel.mash <[email protected]>

[mORMot] using a modified libpq that allows to send the PipelineSync …

9a7037d

…command without flushing prev. commands to socket

github-actions bot added the PR: Please Update label Mar 21, 2023

[mORMot] fix dockerfile, speed up a docker build command

9f08450

pavelmash changed the title ~~[mORMot] send all pipelining comands by one socket send syscall~~ [mORMot] speen up pipelining; speed up docker build Mar 21, 2023

pavelmash changed the title ~~[mORMot] speen up pipelining; speed up docker build~~ [mORMot] speed up pipelining; speed up docker build Mar 21, 2023

[mORMot] shorten server name for better plaintext performance

5832937

sebastienros reviewed Mar 22, 2023

View reviewed changes

pavel.mash added 2 commits March 23, 2023 10:54

[mORMot] change Server header value to mmt because M provokes d…

658165b

…ebate

[mORMot] use 1 letter server header, till is NOT added as a rule and …

1a48475

…verification

NateBrady23 merged commit 9f43d30 into TechEmpower:master Mar 28, 2023

[mORMot] speed up pipelining; speed up docker build #8057

[mORMot] speed up pipelining; speed up docker build #8057

Uh oh!

Conversation

pavelmash commented Mar 21, 2023

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

joanhey Mar 22, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

remittor Mar 24, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

synopse Mar 25, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

joanhey Mar 25, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

joanhey Mar 22, 2023 •

edited

Loading

remittor Mar 24, 2023 •

edited

Loading

synopse Mar 25, 2023 •

edited

Loading

joanhey Mar 25, 2023 •

edited

Loading