freenode/#clasp - IRC Chatlog
Search
1:18:05
cracauer
It is an interesting step. Normally the image saving of a system is done when they have their own GC and the GC is very primitive.
1:19:17
drmeister
I read about a proposed summer of code project that had as its goal to add MPS to sbcl.
1:24:15
cracauer
It was the phase where I was disappointed in GC speedups. All those hacks we tried didn't do much. Following a few pointers more or less (like 2x) didn't really make a dent into the time that was dominated by memory bandwidth and locality.
1:25:50
cracauer
When scavenging (scanning) a block of heap words it pretty much doesn't matter how many words you can skip looking at (in that block). As long as you don't needlessly follow more pointers the CPU just blasts through that. It is in a L1 cache line anyway and yo can noodle around almost for free.
1:29:49
cracauer
What you need is tools that allow you to figure out what could be effective without you having to implement a full GC change.
2:55:52
Kevslinger
I will be entering my third year of undergraduate study in the fall. Though the intermediate steps are unclear (Temple has some programs I may participate in which could extend my time there), the end goal is to enter a PhD program in some form of CS (right now Machine Learning / Natural Language Processing are the frontrunners for concentration).
3:05:21
beach
What cracauer is saying about GC is interesting. I don't understand the details because I don't know what SBCL does sufficiently well, but it will be interesting to see the performance of my planned SICL GC in comparison. Not imminent of course.
3:07:19
beach
It is also interesting to see what the plan for Clasp is. It is good that dead code will be possible to collect, but is live code going to be copied? I seem to remember some negative impact on cache performance for such things.
3:25:06
beach
This pool stuff sounds magic, but in the end you only have a linear memory, so the pool organization is going to have an influence as well I would imagine.
3:27:18
beach
Maybe one can now have any number of "holes" in the available memory and therefore manage any number of independent pools that can grow or shrink.