About TODO Blog RSS Old blog Projects Gallery Notes

Thu, 06 Sep 2007

Containers talk by Eric Biederman and Paul Menage.

Overall notes highlighted in this talk were overview of the support in linux kernel, splitted into namespaces: network, uid, pid, ipc, most of which is already completed and in the kernel, maybe except network, since there is no David Miller's approval on the whole idea. Namespaces TODO list includes additional features (time, dev and other virtual namespaces), isolation issues for physical devices, consolem printks and so on. Next control groups (task containers) were described: memory controller, i.e. limit usage for virtual server, CPU usage limiting, cpusets, task freezer and various per-container limits were noted. Another simple alternatives were shown like rlimits and existing groupping tools (uid/pgrp and others), and theirs limitations. Overhead of the containers was shown to be negligible, but it is possible under some circumstances to trade performance for QoS for some resource controllers. Control group TODO list include virtual servr support and more tight integration with namespaces.

/devel/other :: Link / Comments (0)

Kernel scalability talk by Matt Mackall and Nick Piggin.

Lockless page cache patch and its complexity were highlighted, SLAB/SLUB allocators were presented with short discussion about performance of both (I failed to distinguish SLAB from SLUB pronuncation, but one of them was blaimed as having worse performance, likely it is SLUB, so Matt requested a more precise benchamrk results and testing environment explaination). Number of cache issues and possible solution from avoiding cache misses, moving data between CPUs, MSI interrupt usage in network were shortly concerned by the audience.

There was a note from Christoph hellwig about need for faster block remapping algo used in block layer and short discussion about extent/block based remapping patches.

Problems with page flushing problems and possible ways to fix them were highlighted next. Other issues concerned are larger page usage, pago lock control from filesystems, extent based map, buffer head problems. Scaling down to embedded world showed number of problems with kernel size.

/devel/other :: Link / Comments (0)

Syslet/threadlet and realtime kernel talk from Ingo and Thomas.

Main idea beind -rt patchset is to make system more predictible, not exactly faster in some environments, but more predictible. Ingo highlighted that quite big part of his -rt kernel patchset moved into the kernel already, and that some other bits are pending. Number of features (like irq processing threads) likely will not be moved into the kernel because of its very special nature, although number of its positive sides for existing kernel was highlighted.

Threadlets were shown as a AIO successor, since it is much more easier to program than existing AIO models and any other generic state machines. Its design, implementation, usage cases and possible challenges and problematic places were shortly described.

/devel/other :: Link / Comments (0)

Linux and enterprises.

First enterprise presentation was from DreamWorks. They have more than 600 linux desktops and even higher number of linux machines in the rendering farm. They run own application both batching and iteractive and way how Linux memory management works is not always the best one. Swapping was the main problematic issue with short coverage of swap prefetch patches (each time this words were said the whole audience exploded :). Another issue is OOM condition and oom-killer, which usually kills not what is supposed to be, but it is quite small problem comparing to swap troubles. The last problem discussed is NFS mounting and attributes issues, discussion showed that problem is being solved and there are patches flying around to fix the issue.

Next enterprise client was Credit Suisse bank. This guys want to talk with kernel community directly, since work with distros usually turns out to be completely unproductive. Discussed issues include kernel scheduler, real-time linux ( it was showed that this patchset greatly improves theirs performance without any special tuning), linux diagnostics (some kind of SystemTap, uTRACE and other solutions), TCP/IP issues (Nagle algorithm jitter, various new features of the protocol, like quickacks and slowstart changes). Number of desired features was highlighted related to linux networking like message passing protocols, RDMA and tcp/ip offloading usage, small performance tests of the different technologies and NICs (with and without offloading features) was shown. This rised quite unhealthy discussion with pointing possible bugs in the presentation data, with possibly unconfigured things on the testing side. It was shown that having test setup handy will result in the immediate fix of the problem, but having such testing environment is not always possible.

The last enterprise presentation was from Linux Foundation. First virtualization was concerned with highlighted ideas of sponsoring community developers and providing some kind of management and workgroup for controlling the whole process (which was pointed by audience as a worthless idea). Next issue discussed was power management and so called 'green linux', it was pointed that some areas are not yet fully covered like embedded and servers, and fair amount of works needs to be done pushing hardware vendors. Driver development presentation showed a desire to have open source drivers (surprise), promoting NDA program to allow developers to have an access to hardware, and some kind of business model to work with open source drivers. Presentation briefly showed security issues (like encrypted filesystems and usability of existing features like selinux), scalability (NUMA features, FS and IO scalability isses were described), kernel debugger (it was shown the possibility to sponsor a developer to cleanup and submit KGDB), IPv6 (mostly requirements to make a sertification and meet telco market). Testing projects and efforts in automative process wre highlighted. ZFS easy administration was shown as example of a good taste. SystemTap market readiness was shown, but audience (especially Christoph Hellwig) replied very cold - code is not even remotely in the mergeable form, existense number of wrong design issues in DTrace and SystemTap was highlighted, so essentially probability of its merge is quite low. The last issue discussed is RDMA hooks and need for some kind of generic stack for this technology, which is a bit frowned upon in the community because of it exists like a black box without any ability to debug and fix bugs there.

/devel/other :: Link / Comments (0)

I had a short note about kernel development process.

And my view of the problem of the absence of reviwers and feedback. My point is that when people, which were put into copy list, do not provide any kind feedback, this is just a disrespect for the developer. To fix such hard demotivation I proposed an idea of always answering mails, i.e. if someone is added to Cc: list, then it have to reply. It can be different: 'I do not care, please drop me from copy list', 'I have not time for review', 'I do not like this because of ...', 'This sucks', 'I like it but' and so on. There is no need to answer to every mail, just to show that you got the message, understood idea (if wanted) and provided a feedback. Absence of such feedback really demotivates developers, and thus they will not provide a review for another work, so we have a closed circle: core kernel hackers do not provide feedback to other developers, which do not review other's work.
Idea was met quite cold - no one likes to reply to mails they do not care about, but I think people should start understanding, that until they are changed itself and provide a feedback, others will not move, and thus call from core developers for more reviews is just a hypocrisy.

/devel/other :: Link / Comments (4)