Misc #10628: Peformance of URI module

• Status: Closed
• Assignee: naruse (Yui NARUSE)

[tgxworld (Guo Xiang Tan)]

Please view attached screenshot or go to the following link to see benchmark graph over time.

It got slower after this commit.

Hope this helps.

[normalperson (Eric Wong)]

tgx_world: Please don't use screenshots when a few words explaining the
situation will do. Not everyone can view pics: lack of working GUI,
poor eyesight or danger of opening inappropriate images in public areas.

Fortunately, I was able to open the link to see the issue was with
the bm_app_uri benchmark being slower.

I tried some optimizations, but it's still slower than 2.1.5 on my
Haswell E3-1230 v3 server:

http://80x24.org/spew/m/[e-mail]

Unfortunately, these make evaluating effects of any future compile.c
optimizations more difficult.

benchmark results:
minimum results in each 10 measurements.
Execution time (sec)
name 2.1.5 trunk built
app_uri 0.484 0.584 0.500

Speedup ratio: compare with the result of `2.1.5' (greater is better)
name trunk built
app_uri 0.829 0.968

naruse: any thoughts?

[tgxworld (Guo Xiang Tan)]

Eric Wong wrote:

tgx_world: Please don't use screenshots when a few words explaining the
situation will do. Not everyone can view pics: lack of working GUI,
poor eyesight or danger of opening inappropriate images in public areas.

Fortunately, I was able to open the link to see the issue was with
the bm_app_uri benchmark being slower.

I tried some optimizations, but it's still slower than 2.1.5 on my
Haswell E3-1230 v3 server:

http://80x24.org/spew/m/[e-mail]

Unfortunately, these make evaluating effects of any future compile.c
optimizations more difficult.

benchmark results:
minimum results in each 10 measurements.
Execution time (sec)
name 2.1.5 trunk built
app_uri 0.484 0.584 0.500

Speedup ratio: compare with the result of `2.1.5' (greater is better)
name trunk built
app_uri 0.829 0.968

naruse: any thoughts?

Opps got it. I'll try to explain better in words next time.

[naruse (Yui NARUSE)]

The benchmark consists about two parts: split and new.

In split, the difference of its regexp seems the cause of slow down (for example backtracks of userinfo.
It can be optimized.

In new, some checks are always enabled; it is feature.

require 'uri'

100_000.times{
  scheme, userinfo, host, port,
    registry, path, opaque, query, fragment = URI.split('http://www.ruby-lang.org')
  URI::HTTP.new(scheme, userinfo, host, port,
                                     registry, path, opaque, query,
                                     fragment, URI,false)
}

[naruse (Yui NARUSE)]

Anyway I'm ok for Eric's patch.
Could you commit it?

We should consider ext/uri or strict/loose API for Ruby 2.3.

[normalperson (Eric Wong)]

[e-mail] wrote:

Anyway I'm ok for Eric's patch.
Could you commit it?

r48980

We should consider ext/uri or strict/loose API for Ruby 2.3.

We should be able to improve compile.c to do more optimizations.
I prefer to avoid any new C-exts.

Things like [Feature #10423] (reduce string literal allocations).
Also, this patch gave me a new ideas: we should be able to merge/share
nearby call_info structs on the same object. For example:

m["a"], m["b"], m["c"]

May all use the same call_info object, I think:

{:mid=>:[], :flag=>256, :blockptr=>nil, :orig_argc=>1},

This can improve IC hit rates and reduce memory.

[tgxworld (Guo Xiang Tan)]

Latest results so far after the patch :)

https://railsbench.herokuapp.com/tgxworld/ruby?utf8=%E2%9C%93&result_types%5B%5D=app_uri&commit=Submit

Do let me know if you need any more information. Happy holidays!

[normalperson (Eric Wong)]

[e-mail] wrote:

Latest results so far after the patch :)

https://railsbench.herokuapp.com/tgxworld/ruby?utf8=%E2%9C%93&result_types%5B%5D=app_uri&commit=Submit

Do let me know if you need any more information. Happy holidays!

I actually can't see results in my text-only browser, just that
bm_app_uri is selected. Can you just show us numbers? Thanks.

[tgxworld (Guo Xiang Tan)]

Eric Wong wrote:

[e-mail] wrote:

Latest results so far after the patch :)

https://railsbench.herokuapp.com/tgxworld/ruby?utf8=%E2%9C%93&result_types%5B%5D=app_uri&commit=Submit

Do let me know if you need any more information. Happy holidays!

I actually can't see results in my text-only browser, just that
bm_app_uri is selected. Can you just show us numbers? Thanks.

On my builder, these are the following results:
before regression: 0.97
regression: 1.1 ~ 1.2
after patch: 1.0

[hsbt (Hiroshi SHIBATA)]

• Status changed from Open to Assigned

[hsbt (Hiroshi SHIBATA)]

• Status changed from Assigned to Closed