Forum OpenACS Q&A: Request Processor Performance

Collapse
Posted by James Thornton on
Is anyone working on improving the performance of the request processor?

We were going to use OpenACS 4.5 for http://bloglog.net, but benchmarks showed that we are only getting ~8 DB PPS whereas we are getting 20+ DB PPS with OpenACS 3.5.

Is permission checking the primary bottleneck? Is there a way to disable permission checking for a page, directory, or file type?

Thoughts?

Collapse
Posted by Roberto Mello on
Do both pages do the same thing? Same number of queries? Same permissions checking?

Don's permissions, which should make it into 4.6.1, should help quite a bit.

And why are you looking into 4.5 when 4.6 is out? 4.6 is in much better shape.

-Roberto

Collapse
Posted by James Thornton on
<blockquote> Do both pages do the same thing? Same number of queries?
Same permissions checking?
</blockquote>

Both pages execute the same query (only one query), but the OpenACS 3.5 page doesn't perform permission checking (which is what we want for this page). Ideally, I want to turn permission checking off for certain pages -- has this feature been developed for 4.6?

Collapse
Posted by Peter Marklund on
James,
experience from tuning my homepage indicates that you can cut response times a lot by turning on caching for permissions (an APM parameter of the kernel). You should mount developer support, activate query logging (APM parameter of the package) and restart the server. Figure out which queries are the bottlenecks and consider caching with util_memoize, see this thread:

https://openacs.org/forums/message-view?message_id=77562

Recently I had been looking into unneccessary queries that the request processor is making. One such query that you might see is

        select node_id
    from host_node_map
    where host = :host

- not exactly an expensive query but still, it should be cached. The query will be executed if the hostname in your AOLServer config file doesn't match exactly the host header of the request. For my homepage that means the query is executed for marklundweb.com but not for www.marklundweb.com (what I have in the config file).

When running developer support you may see a query (executed twice sometimes) building up the context bar. This query has been replaced with a cached lookup recently. I think this change will be in 4.6.

On the OpenACS development branch (what will be 4.7 eventually) there is an overhead query that is run four times per request to check the locale preference for a user for the requested package. I'm working on adding a cache to elliminate those queries.

Another thing to do is set the performance mode kernel APM parameter to 1. This will reduce the amount of code executed in the request processor on each request.

What is PPS? Are you using Oracle or PG? What kind of hardware are you on?

Collapse
Posted by Peter Marklund on
James,
all unnecessary over head queries invoked on every request that we have found so far have been cached now (including the ones mentioned above such as host-node-map and user locale lookups on 4.7).

If there are any that we've missed, let us know and we'll squash them.

/Peter

Collapse
Posted by Don Baccus on
As Peter mentions set performance mode to "true".  This causes it to map from the URL directly to a file.  You lose immediate tracking of (say) replacing a .adp file with a .tcl file but get speed instead.  It also disables the check for "watched" files.  So ... for development set performance mode to "false" but only then.

Also I thought I'd added my site node caching of permissions work I did for Greenpeace to OpenACS but apparently I forgot to do so.  I will do this soon.  It tracks changes to site nodes accurately so if you shut a node off to the outside world by removing the "read" permission it will be honored immediately.

General caching of permissions is a *bad* idea, IMO.  Making permissions fast enough to use combined with higher level caching where there's more control is much safer I think.

The new permissions implementation is almost complete for PG, BTW, so there's hope there.

Collapse
Posted by Don Baccus on
Does the permission caching stuff time out?  It looks like it calls util_memoize without a time out.  This would be *very* dangerous ... the cache can become incorrect if code changes an object so it doesn't inherit permissions or gets them from a new context, etc.
Collapse
Posted by Andrew Piskorski on
The OpenACS kernel (at least as of OpenACS 4.6.2) has a PermissionCacheP switch. Above, Peter and Don mentioned this back in February, and Don at least was raising potentially serious concerns of whether the implementation of PermissionCacheP was correct or even a good idea at all.

So before I start digging into the PermissionCacheP code myself, what was the conclusion of that discussion? Is PermissionCacheP now robust and a good thing, or are there still problems or concerns?

Here are a couple other threads related to caching of permissions queries, from Feb. 2003 and June 2002

Collapse
Posted by Jeff Davis on
The permissions are cached just fine and properly flushed when updated in most cases (if you use the API rather than hitting the db directly that is). The problem is that if you add or remove someone from a group (or add or remove a permission on an object for a group) the cached value can then be invalid. This is the case for pretty much any inherited permission.

Properly maintaining the cache in that circumstance is very expensive and consequently no attempt is made to do it.

So the short answer is that if you have permissions set up in a particular way on your site it *might* work but but generally there are a lot of cases where the cache is out of sync with the DB. With Don's permissions checking speedup work it is hopefully not necessary for most sites anyway.