LSE Conference Call Minutes 11/16/01

*- Indicates Action Item

I. Hubertus Franke-

	Read the proposal and the ALS paper. 

	Davide Libenzi wrote a patch that records number of jiffies a task
	accumulates between the time it is in and out of the runqueue?

	push/poll model of load balancing.
	poll- the idle process looks for tasks to take, so it never really
		goes in standby. Done by Davide Libenzi.

	push- the idle process is awoke with an ipi when a task wants to run on it. 
		enables proc to go in standby. 

	MultiQueue scheduler has a problem at the low end with load <=1. 
	So there davides patch could make some impact.  Anything that improves 
	the low end will help us.

Gerrit*- Invite Davide Libenzi to IBM Beaverton to hash some of this stuff out.	

 	Numa support may be put in separately.

	Steven Tweedie wants to see a wider range of benchmarks (bonnie mentioned).

	Tim Witham- the osdl with stp is all set up do this the only
	thing we are missing is the actual generation of the tests.
	
	ruth- try polygraph. a web cache workload.

	Dont worry about 8 way performance until you get the low end performance better

II. Tim Witham

	The STP is an automated system to test your patches. You pick the size of 
	system (1-16) then the os is installed with your patch and the test is run 
	with the results logged, all done automatically.

	We have stp up and running. I encourage everyone to go and look 
	at it: http://www.osdl.org/stp

	Looking into cerberos.

	limited to lilo options right now.

	Q- is it safe to assume the base kernel has run through these tests and passed?
	A- yes, you can look all this up on the web.

	Q- how long does it take to clean and reinstall the kernel with patch?
	A- it is done asynchronously, so it doesnt impact your testing time directly.
	a complete reinstall is done in between every new patch. Looking into
	doing a quick loop but not set up yet.


III. Paul Dorwin- 

	System topology. Provides infrastructure to discover various hardware
	components within a system. Currently working on patch for i386 which 
	will utilize mps table to populate the kernel data. Next will work on 
	ia64 patch. Goal is to provide patch for various platforms.

	By the next NUMA meeting hopes to complete api to include both kernel and 
	user space to make decisions using this api.

	Will put more api info in the document and convert it to html to put it on
	the lse/numa web site.
	
IV. Paul Jackson

	Another update on processor memory placement, cpumemsets. It looks like 
	there is a good overlap with paul dorwins work. 

	The changes made over last few weeks:
		1. anybody can mess with maps but only root can add resources.
		2.simplify the memory layout you get back from the query. 	
		3. settled on the name memory blocks for chunks of continuous memory.	

	Hoping to get first draft patch out by the end of this month, although a 
	little behind schedule right now.

V. ALS

	Papers online at:

	http://www.linuxshowcase.org/tech.html

	No one had anything to say about it.

	
VI. Mingming Cao

	Working on diskio statistics scalability. 
	
	Current implementation uses static array indexed by major number and
	 disk part of minor (less than 16). Limitation is you can only do 
	statistics for major devices less than 16.

	Two proposalis to solve this problem:

		1. Change array from static to dynamic
		2. Move statisticis info to request queue. 
		the lookup is much faster.  Also self contained data 
		structure so when do make request already in request 
		queue. Dont need to look up major and minor number. 
		Works for both scsi and ide.
		Problem is this is not diskio statistics but is request queue. 
		Cant always tell which disk is doing real io because device 
		driver could make its own request queue, two disks could use 
		the same request quewue.

	Q- what could cause you to put two disks on one request queue?
	A- as far as I know ide could use 4 disks in theory but have only 2 queues.
		currently onlys uses 2 disks.  this is all defined in driver code.
		if only ide problem could try to fix there, but dont know for sure
		if other drivers do the same thing.

	Q-How many peopole would want to see the statistics along the partition level
	vs the disk level?

	Q- would it be possible to make this configurable  whether they want
	statistics at partition or disk level.
	A- yes that is possible.

	Q- What about LVM?
	Suparna- LVM remaps a request, doesn't have a request queue of its own. 

	Mingming- it has a request queue, just doesnt use it.

Mingming*- will put patches and proposals on the lse web site.

	Q- How is the performance?
	A- Proposal 2 should be really good for performance, havent tested yet.

	sounds like a canididate for the STP.