[Ur] TechEmpower Benchmarks

[Adam Chlipala]

On 12/12/2013 01:50 AM, escalier at riseup.net wrote:
>> Again, I'd be very interested in any help anyone is willing to offer on
>> comparing the execution of the latest Ur/Web benchmark against one of
>> the current winners for the 'plaintext' benchmark.
>>      
> A sketch of that based on limited testing:
>
> Vert.x is the current 'plaintext' winner and 3rd for 'json'. When I ran it
> against Ur/Web I noticed some extraneous headers in Ur/Web's responses.
> There was a duplicate 'Content-length' header (there was already a
> 'Content-Length' header) and a 'Content-script-type' header that was doing
> nothing in that context.
>    

I'm surprised that this change made such a big difference!  Are clients 
connecting from the same host where the server is running?  I've only 
tested in that setting, and maybe a bit of extra header noise is 
substantially costlier when it actually needs to transit the network.

I've pushed changes to remove the duplicate/unnecessary headers you 
point out (not quite using your patch), but (in localhost-to-localhost 
testing) I don't see any clear performance change.

> Removing these bumped 'json' performance from
> `621613 requests in 10.00s, 123.90MB read`
> to
> `651695 requests in 10.00s, 93.85MB read`
> . This puts Ur/Web quite close to JITted Vert.x at
> `672104 requests in 10.00s, 97.43MB read`
> .
>
> On plaintext, I got Ur/Web at
> `679445 requests in 10.00s, 84.88MB read`
> and warmed-up Vert.x at
> `715567 requests in 10.00s, 90.08MB read`
> .
>    

I imagine performance shouldn't be worse with the latest Ur/Web version, 
so that leaves just a small puzzle of what accounts for Vert.x's 
plaintext advantage.  Thanks for running this comparison!