List of changes:
1. Created a separate thread pool for handling cron and ping tasks. The size of the pool is based on direct.agent.pool.size. The existing direct agent pool will run all commands other than cron and ping.
2. For normal tasks (generated as part of user/admin API calls), if throttle limit is reached then tasks get queued up for subsequent execution once threads are available.
3. For cron and ping tasks (internally generated by MS like ping, VM sync etc.), if throttle limit is reached then these gets rejected. Since these are internally generated these can be rejected without any issues.
Cloudstack sends requests to directly managed HV hosts (direct agents) using the direct agent thread pool. The size of the pool is determined by global config direct.agent.pool.size defaulted to 500.
Currently there is no restriction on the number of threads a direct agent can use from this shared thread pool to send requests to the host. This is fine as long as the host is responding to requests
in a reasonable amount of time. But if there is a considerable delay in getting response, the thread remain blocked for that much time. As more commands are send to the slow host threads keep getting
blocked. This can eventually lead to a situation where requests to healthy hosts cannot be processed as there are not enough free threads.
The problem being addressed here is to localize the impact of few bad hosts, so that entire management server is not affected.
One such way is to throttle based on the # of outstanding requests on per host basis. The outstanding requests to a host will be a % of direct agent pool size. This is configurable based on
direct.agent.thread.cap. The default value is 0.1 or 10%, a value of 1 would mean the old behavior where there is no upper cap. This will ensure that the impacted host will be bound by a upper cap on the number of threads it can use to process requests and not the entire pool.
was discussed on the mailing list as a useful debugging tool, currently
the log prints the DB id of the agent, which makes admins have to look
it up to know where the Command was run.
The managed context framework provides a simple way to add logic
to ACS at the various entry points of the system. As threads are
launched and ran listeners can be registered for onEntry or onLeave
of the managed context. This framework will be used specifically
to handle DB transaction checking and setting up the CallContext.
This framework is need to transition away from ACS custom AOP to
Spring AOP.
Santhosh Edukulla
Koushik Das
Anthony Xu
wrodrigues
Daan Hoogland
Alex Huang
Marcus Sorensen
Darren Shepherd