peterd's blog
The RI seems mostly happy with Linux version 2.6.26.3
Fedora 8 just updated from 2.6.25.something to 2.6.26.3. I ran the TCK on my current development version of the RTSJ RI without recompiling it under the updated system and it worked fine except that the physical memory classes are all failing when they try to mmap the requested memory. It looks like the kernel developers think this is a feature. I hope I don't have to restrict the kernel to Linux prior to 2.6.26.
System description language
The analysis applet uses a simple system description language. Each line represents a task. It contains the name, period, cost, deadline, blocking time and priority for the task. It's simple for a human to enter and simple for the applet to parse. It also fits nicely with what RMA likes to analyze. However, it only describes simple systems, it forces the user to compute the blocking time for each task, and it begs for a pre-processor.
Soft real-time analysis
I've run into an interesting dissertation (warning -- large PDF document) on probabilistic analysis of soft real-time systems. Being from 1999, it's a bit old, but it seems more useful than more recent things I've found.
More changes to the beta analysis applet
When I added the code to sort the results table on any reasonable column, I noticed that the priority column sorted the tasks backward. Hmm. So I fixed that.
Surprisingly, the analysis code did not totally explode, though some results were a bit strange. (Most analysis code assumes that the tasks are sorted so the higher priority tasks are at lower indexes.) I fixed that. Now sorting the output table and sorting for analysis use different sort methods.
Latest on the Beta Applet
Today I made two changes to the applet:
- Clicking on the header for a column will sort the results table on that column. No, clicking on it twice won't sort it in reverse order. :-)
- There's now an exponential time algorithm for finding a feasible set of priority assignments for MP systems. The applet will try this algorithm if the Audsley's algorithm (heuristic) fails and there are no more than 16 tasks.
I wonder if non-RM/DM priority assignment is really better for MP
The example in my previous post is, I think, typical of the problems used to show the nastyness of scheduling for a multi-processor. It has three tasks:
- t1 uses 80% of a processor with a period/deadline of 100.
- t2 uses 50% of a processor with a period/deadline of 40.
- t3 uses 25% of a processor with a period/deadline of 40.
Multi-Processor Response Time Analysis
The beta analysis applet now supports the multi-processor response time analysis I've been writing about. To me, a graph of response times gives a better feel for the system than simple feasible/infeasible answers.
Deadline monotonic priority assignment doesn't seem to "work" for multi-processor, so I'm trying Audsley's algorithm on MP. I haven't thought about it enough to know if it will always find a feasible priority assignment if one exists, but it does seem to work.
Response Time Analysis for MP systems!
Ted Baker (a CS professor at FSU) suggested I look at a paper by Bertogna, and Cirinei from the 2007 RTSS that pushes the state of the art for RT analysis of MP systems. I can't find a fully-accessible version of the paper on-line but here's the ACM portal page for the paper. I'm a bit excited about this new analysis algorithm.
The RI download page has new stuff
It has version 5 of the Alpha RI for RTSJ version 1.1. The major 1.1-related changes are a re-designed API for phasing control and an API that supports setting processor affinity for Threads (Java, RT and NHRT) and bound async event handlers. It also has default affinities that apply when it doesn't make sense to inherit affinity and for unbound async event handlers.
It also has version 6 of the 1.0.2 RI. This is a bug fix release. The main bug fixes have to do with the interactions between ATC and locking.
