This part of this post assumes that the reader has read my last post.
This post contains embarassing news: In my effort to organize the patches to make it easier to run a script that benchmarks each change to see how it affects performance, I made a very simple mistake, which mostly would disable Immix. This was done in the first patch of the stack, so all patches were compromised. I noticed this while I was attending Simon Marlow’s request of measuring how many lines are being freed and allocated, and notice that it was none. My error was simply to set two different constants with the same value: BF_MEDIUM and BF_SWEPT. After the correction, I noticed two more errors that were introduced in this period, and that were not corrected because, as Immix was not activated, they were not causing any harm. Now, with everything in the stage I was before: a memory corruption that happens ocasionaly only in the nofib/real/maillist benchmark. As described in this link, I’ve tried already a lot of different techniques to debug this one, but it’s being hard. In the mean time, I’m measuring again the stack of patches, in a lab using the complete nofib (except for nofib/real/maillist) and nofib/gc, and in my laptop only fibheaps, fulsom and constraints, from nofib/gc, which were used in my last post.
Debconf was great. I’ve stayed in New York with my family doing tourism for some days before it, and New York city is fantastic. A very diverse architecture and people from all around the world. It’s really impressive. I arrived in DebCamp on Friday and it was not easy for me to find where to check-in, and where to go after unpacking my stuff. I assume I haven’t paid much attention to the e-mails explaining this, but I tried to search them when I got there, without success. It was good, on the other hand, to walk on Columbia searching: it’s a very nice campus, not comparable to anything we have here in Brazil. During DebConf I tried to help the organization as much as I could, being Talk-meister in some talks, helping a little bit with food organization and participating in the video team. All of this was very fun. There were some talks that I couldn’t attend to, and I watched the videos, or part of them, here on Monday: Making Debian rule again, FreeBSD, Anti-features… I still want to watch some, like the Analysis of Debian mailling lists using machine learning, and there are others that I wanted to see but are not recorded, like the one about notmuch. I got my key signed by a lot of people, which certainly was not the case before, and it the GPG best practices BoF was great. And there was the proposal to host DebConf12 in Brazil, maybe in my city, Belo Horizonte. Which reminds me: