Lavonardo

Registered to the conference. Good loot - not one, but two t-shirts (the other provided by the newly minted Intel Open Source Labs). Bought a hardcopy of the proceedings as well. By-invitation annual kernel summit and the very first desktop developers’ conference were ongoing, took a peek into the latter (based on a six second analysis, the guys were discussing the future of KDevelop).

The adjoining Rideau Centre mall is its old trusty self. Food court with variety and a smallish selection of interesting shops. HMV’s sales are as good as ever. EB’s PC game selection has gone down some 80% from last year’s. Radio Shacks have mutated into something called “the Source”.

Annual Ottawa Linux Symposium is upon us again. And this time the proceedings were published on the official page before the actual conference.

Not much different from previous years, standard seems to be high and there are a lot of familiar faces and names among the speakers. Yet another good crop. And this continues to prove the fact that OLS is, indeed, the technical conference on Linux.

Conference finished now - four days of pretty intense scheduling of presentations. Today was a little shorter than the previous.

Creating Cross-compile-friendly Software by Sam Robb. A presentation by a Timesys employee describing the wacky world of cross-compilation (no other term used even when the architectures are the same). Differentiation between build / host / target systems.

I would hate user-space locking… by Iñaky Perez-Gonzalez. Robust mutex-work (dying threads no longer abandon held mutexes) and its effects on the threading library. Presentation available at project page at OSDL. Needs gluecode to make NPTL to use the described lock mechanism, struggle is still ongoing, a definite chicked & egg problem once again, between kernel/distribution vendors/glibc maintainers. Provides previously missing POSIX margin codes, and thus enhances Linux’s position as a porting target.

BoF: CE Linux Forum Introduction led by Tim R. Bird. Founded 2003 by Sony & Panasonic. Main focus on kernel enhancements thus far. Six subgroups: bootup time (most active thus far), system size, power management, real time, audio/video/graphics and security. Attempts to escape specification-fever by calling documents ToDo-lists. Call for a sensible communication forum - multiple exist, but none are optimal: lkml, arch-specific lists, handhelds.org, embeddedlinux-book’s community.

BoF: udev led by Greg Kroah-Hartman. Udev is shipped in multiple distributions (SLES9 the biggest includer). Public expects a naming scheme, now that the mechanism is done. OSDL has picked up the ball on this, with SuSE providing the initial scheme (list). Lots of unnecessary flak from a talkative Red Hat engineer. Half a dozen hardcoded devices, but almost everything now discoverable via sysFS.

Andrew Morton had a good keynote (though not as funny as Rusty Russell’s last year). Time for dinner. And the official dead dog party.

EDIT 5.8.2008: OSDL links gone stale.

Comparing and evaluating epoll(), select() and poll() by Louay Gammo. Turns out it’s actually someone else talking. Using a 1-byte payload and a very simple HTTP-server, the event-handling mechanisms are compared. Turns out that there’s more gains in application code (optimize handling of incoming connections), rather than in kernel space. Most alternatives (poll, select, epoll, used in level- and edge-triggered fashion, vectored epoll) do not differ in a significant enough fashion.

new development model / kobjects & krefs by Greg Kroah-Hartmann. Two presentations for the price of one. First dealing with the “new development model”. Meaning that 2.7 will not be opened for a long time, and that 2.6 will keep accepting new features. This is how things have been working lately, and they have indeed been working well. There’s a good description of the process at lwn.net, as well as plenty of discussion on lkml and news sites (such as kerneltrap.org). Slides available at Greg’s homepage.

The other topic, reference counting in kernel data structures was far more technical, and shorter. Still uses Big Kernel Lock, but that’s not seen to be a major issue, even with regard to performance. Slides available at Greg’s homepage.

Xen and the Art of Virtualization by Ian Pratt. Lots of good material available at the project’s home page. Xen has advanced with leaps and bounds since the last peek inside, and especially performance-wise seems to be a very credible implementation indeed. It uses the hardware protection features of the x86-CPU to a large extent to provide trusted hypervisor environment in which the “guest operating systems” (Linux, BSD variants, Windows XP) run (in a slightly less priviledged mode). The biggest omission right now is the SMP-incapability of the guest operating systems. Implemented as a separate architecture in the source tree to keep the code clean.

TIPC: Providing Communication for Linux Clusters by Jon Maloy. Another “no big news” presentation. Ericsson’s TIPC has evolved a lot, and nowadays even looks like kernel code. On surface looks like a good TCP/UDP replacement: connectionless/connection-oriented, reliable/unreliable, single node/cluster. Until the “not routable” feature comes into play. TIPC is usable in small environments (ie. perfect for intra-cluster messaging), but in the big picture it has problems.

The World of Open Office by Michael Meeks. Rather annoying presentation style. Ingratiating, yet at times very interesting. Open office has grown up. And it has also grown big. The building process is scary and has consumed many spirited hackers. OO is very C++-version specific, and used to be shipped with a lot of redundancy in the libraries to provide the correct ones in the environment used. Nowadays things are better.

BoF: Linux Kernel Scalability and Tradeoffs led by Dipankar Sarma. Yet more IBM-led birds of a feather sessions. Mainly dealing with locking issues (a presentation in the following morning by the same person). Rapidly descends into the very precise questions (is RCU good for soft IRQs or not).

BoF: RAS: Efficient Data Transfer led by Richard J. Moore. RelayFS, a simple method for shipping vast amounts of data in/out the kernel is not in 2.6. Has a chicken & egg problem, and needs good stories to convince the powers that be to include it. Key reason for not including is its similarity with netlink. Which does bear similarity, but couple of key features (zero copy) are missing from the latter. Would benefit from true benchmarking between the two. Also, providing rFS as a loadable module would give people ability to mess with it without patching their kernel.

Improving Linux Resource Control with CKRM by Rik van Riel. One of the potential additions to 2.6. A very interesting project from a high availability point of view. Semi-equivalent features exist in other UNIX-implementations. Good material exists on project’s homepage.

Linux Virtualization by Chris Wright. An introduction to virtualization as an operating system feature, and a recap of the existing implementations. Rather academic and high-flying, but decent nonetheless. Multiple references to the presentation on Xen. Hadn’t planned on seeing that session, but best laid plans have to change…

Towards Linux-based Open Telecom Platforms by Ibrahim Haddad. A presentation by a colleague from OSDL’s Carrier Grade Linux-group (works for Ericsson). Nothing new here, a basic outline on the reason’s for the CGL-group’s existence, listing of accomplishments (honest, the new spec is coming), and a brief recap of the projects that Ericsson has been working on: AEM, TIPC, DigSig. On the last of which I have a spontaneous feature enhancement, but decide to hold on until I’m back at work. Would be cool to be working on it, but likely will be an extra-curricular effort on my part. A mini CGL-session ensues after the presentation when relevant attendees crowd the poor presenter.

Linux Kernel Hotplug CPU Support by Rusty Russell. An entertaining session that has very little immediate value, since CPU hotpugging is not often done. Basically there’s a framework for doing this already, but it has some key issues. Fiddling with CPU affinity during a hotplug event may thoroughly hose the system. All actions are currently serialized, which may be a problem on truly big iron (which SGI’s already using).

BoF: First Failure Data Capture Using Crash Dumps led by Hariprasad Nellitheertha. Preaching to the choir here… Most of the audience does indeed subscribe to the “post mortem debugging is a must”- theory. IBM’s been working on a kexec-based dump (basically switching to a healthy [streamlined] kernel to take the dump and then back to the production kernel). Some people use Linux Trace Toolkit in “flight-recorder mode” to capture the system’s last moments (trace buffer stored in the dump). Long discussions on what would be the optimal tracing tool - gdb seems to be the least hurtful for most attendees. It needs some work with regard to exception frames. Kexec is nearing completion, but is by no means a complete solution - drivers need to be fully re-initializable to be reliable in a kexec’d environment.

BoF: How to Help a Penguin Get Up Quickly after a Fall? led by Vara Prasad. Spaced out a couple of times during this presentation (long day etc). Yet another IBM-led discussion on internal improvements. kprobes seen as a usful dynamical instrumentation tool, oprofile seems to be headed for a bright future as well. Discussion kinda dithered on many topics (since the title was not exact enough - boot time related discussions were aborted early, taken care of in another session).

Visited Chilly Chiles near the convention center. A small shop that sells nothing but chiles and chile-related products. Grabbed the new catalogue, and bought some salsa and chips for evening use. Turned out to be between wohoo! and hothothot in heat. Which is always good. Definitely have to pay a second visit. My kitchen is empty without a poster of the different peppers of the world. And I need to bring some hot sauces back as well. Though the chips are the among the best I’ve tasted, shipping is a problem (consume vast amount of room in the suitcase).

EDIT 5.8.2008: A few links have gone stale and removed.

Where 2.7 is going by Jonathan Corbet (slides available at: lwn.net. Quite badly sidelined by the “new” working method / numbering scheme. But still an interesting overview of the next years or so.

NFS4 and rpcsec_gss for Linux by J. Bruce Fields. Protocol security not really my cup of tea, here more to see what’s up with NFS4 in general. Server-restart recovery is the key omission in the current (2.6.7-level) implementation. NFS4.1 work has been started (at least on paper, it was not stated whether this was Sun-led or a CITI-effort), the proposed features sound interesting: cluster file system semantics in failovers (with all the needed symmetric state preservation needed), parallel I/O and something called NFS over RDMA. No slides available, but the project homepage packs a lot of information.

TCPfying the Poor Cousins by Arnaldo Carvalho de Melo. An interesting talk about reusing TCP-code in other protocol families to ensure code quality and eliminate special cases from core networking code. DCCP (a new protocol, rather like SCTP) is a proof-of-concept for the process.

Methods to Improve Bootup Time in Linux by Tim R. Bird. First fruits of the Consumer Electronics Linux Forum-initiative. An impressive presentation, that proves that in a constant-hardware environment there are really significant improvements to be gained. The target (having userspace running in 500 milliseconds) is a tough call, but the presenter looked surprisingly hopeful. Execute in Place is helpful, but adds artifacts to delays - partial XIP (hotspots in RAM) a possible solution. There have been multiple .rc script replacements, but true parallelization of scripts is still missing (would need a re-factoring of init)

BoF: Linux Scalability Effort led by Hanna Linder. First birds of a feather-session gets off to a good start by a big bunch of IBM-guys presenting the bottlenecks identified (and in some cases eliminated). dcache, NUMA environments, ckrm discussed - all have issues.

BoF: Hardware error Reporting, Handling, Debugging led by Robert Hentosh. Hardware debuggers thought to be the best way to approach debugging, but that’s only available in lab-environment. Key issue is to provide enough information runtime to resolve any issues in post mortem-fashion. NMI is a totally inadequate solution (it has very little payload), and machine checks are slowly replacing it in modern CPU architectures (which does not help the zillions of installed x86-base at all).

Exchanged the frame-challenged Shrek to Once upon a Time in the West at HMV.

It’s getting hotter. Decided to have dinner at Subway, since the AMD-party’s bound to have snacks.

And the party indeed had food. Of finger kind, but decent nonetheless. A very long-winded presentation having nothing to do with Linux, AMD or pretty much anything else. No-one I knew won anything in the lottery (some quality HW was dished out).

Stopped by Boland’s on the way to the hotel. DNS still acting up, e-mail less existence is not painful at all.

Flew to Ottawa, Canada to attend the annual Ottawa Linux Symposium.
Proceedings have not been publically announced yet, but are available as: vol. 1
and vol. 2.

Flight was far from too short: Helsinki-Stockholm-London-Ottawa. Guess it was a wee bit cheaper with the extra leg.

Layover in Arlanda was a piece of cake, easily accessible Diners Club lounge that pretty much sucks one in…

Heathrow was its horrible old self. Still being renovated. Still hot. Still without air conditioning. Still loud. Ate at a pub-type place (no matter where you eat in Heathrow, the food tastes like pub-grub). Layover was mercifully short.

Flight to Ottawa surprisingly short and pleasant (go, Air Canada, go). Multiple small meals, attentive waitresses. Movie was Starsky & Hutch, missed bits due to eating and due to horrendous sound via less-than optimal earphones.

Luggage arrived without a hitch, the hotel located conveniently close to the congress center and next to Byward Market, a vast conglomeration of bars and restaurants. Had dinner at a “traditional” Mexican place where the menu claims some old Aztecan ways are used in food preparation. Food’s good. Don’t remeber the name of the place.

Off to hotel, where fiddling with the complimentary WiFi-connection’s settings and baseball on television prevent from falling asleep too early.