On Fri, Aug 31, 2012 at 9:05 AM, Ed W
<..hidden..> wrote:
Hi folks, I have spoken to Chris about this a little offlist, so I just
want to stress that this is meant to be a positive sounding question!
What kind of speeds are you currently seeing page render requests with
Plack/Starlet? I just want to jump in with an abstract benchmark, but
we should be aiming to get all screens which do trivial backend
processing to be performing in the sub 100ms response time frame. In
fact for trivial screens on a sensible sized box we really should be
seeing sub 50ms response times.
I have not run through profilers specifically. I haven't quite figured how to do this yet. However, on my system, basically the way it works is that we preload all the expensive modules and then, fork, run, die. This prevents legacy SQL-Ledger code from corrupting the Perl interpreter on multiple runs, which actually happens and it means that SQL-Ledger code is not cached. We could probably cache more LedgerSMB code and speed things up a bit but that would be marginal gains so far and not 100% sure how helpful it will be.
Typically what I have seen on my development system is a drop from approx 3 sec-5 sec response time to most requests to well under a second. I have been unable to hook profilers in and don't know how much of the delay now is due to browser processing, but the speedup is significant. Profiling on the scripts run as full CGI show over 60% of the time being thing like initializing the localization classes. This is all moved to service startup in this model.
I am getting errors on Starman (not Starlet) but not sure if it is my environment or the .psgi or the code or what.
As a data point I am currently working on a mojolicious (perl) web
configuration interface for an embedded (ish) box which runs on a 500Mhz
Geode (Alix). This generates real web pages in around 80-120ms despite
the rather wimpy processor (those are "to the browser" speeds, not just
rendering speed). I haven't benchmarked on a proper machine, but
roughly this machine benchmarks at around 10% of my Q6600 2.4Ghz machine
(single threaded).
I'm not sure if you have looked at Mojolicious, but it's now a very
mature and flexible toolkit for building all kinds of web services. The
code is fantastically clear and concise and well worth a look for
inspiration. Basically all of its features are simple packages and have
been implemented in a remarkable simple fashion. It does seem to steer
a decent line between "opinionated" and "flexible"
I would recommend it as an interesting option for LSMB, not least
because it's more of a "VC" than an "MVC" system, you are left to
implement the "M" yourself. It also provides a lot of the plumbing for
templating systems, ajax, pjax, webservices, etc. It's such an easy
architecture to work with it's likely a potential "drop in" change...
I think we should be further discuss this perhaps on a separate thread (the Mojolicious bit).
Best Wishes,
Chris Travers