Slashdot Mirror
Linux 2.6 Kernel Stability Freeze
An anonymous reader writes "Linux Creator Linus Torvalds released the 2.6.0-test7 Linux development kernel today and declared a "stability freeze". It has been made quite clear that from this point only "strictly necessary stuff" will be accepted, clearing the way for an official 2.6.0 release sooner than later... possibly at the end of this month."
2.5 has been largely successful, and a lot of end users were able to compile it. 2.3? That's another story. I remember not being able to compile 2.3 once.
Good job to all the kernel hackers.
The October of cool new toys:
Sony PSX
Panther (Mac OS 10.3)
2.6 kernal
Half LIfe 2
Ow! Ouch! Sorry!
Oh, I dunno, SCO registration form on the first boot-up?
I wrote a speaker bracelet module. Alas, it's been rejected because I turned it in too late. It was really cool though.
When the 2.4 series came out, it was much criticisd for not having anything near the stability of the old 2.2 series (I'd say it haven't catch up yet,but since I use it in a desktop machine 2.2 is not an option)... What can we wait from the brave new world the 2.6 kernel will bring?
Where is that guy who'd die defending what I had to say when I need him?
Wait! Wait! Wait! I've got a million lines of SCO code I want to insert!
The CB App. What's your 20?
Good question. It will appear in someone's (I believe AA's) tree in a couple months or weeks. From there it will stabilize and get added to 2.6. Here's the latest status update from Hans:
The filesystem is getting reasonably stable.
This weekend we hit a bug in space reservation, which we can't reproduce yet but probably isn't too hard to find by code inspection. There is some thought that the assertion not the space reservation is buggy, in any case we'll release a snapshot after it is fixed.
Our performance is generally wonderful and getting better.
It has the following weakpoints:
* We allocate a "jnode" per unformatted node in the filesystem. The traversing of these jnodes consumes more CPU than performing the memcpy from user space to kernel space when doing large writes. I don't yet really understand on an intuitive level why this is so, which is a reflection on my ignorance as it is consistent with stories I have heard from other implementors of filesystems who found that eliminating per page structures was an important part of optimizing large writes. We will fix this by creating a new structure called an extent-node that will exist on a per extent basis, and this will probably cure the problem. This will greatly simplify parts of our code for reasons I won't go into, and it will also take us 6 weeks to do it. I don't think users should wait for it, and so we will ship without it.
* Our dbench performance was poor, has improved due to coding changes, and we need to test and analyze again. Perhaps more fixes will be needed, we can't say yet.
* Our fsync performance is poor. We will pay attention to this next year, frankly, after we have fully implemented the transactions API. At that point we will say something like, if you care about fsync performance you should be using the transactions API and/or sponsoring us to tune for NVRAM, users will say back "but our legacy apps on hardware without NVRAM matter!", and we will grudgingly but effectively tune for this because we care about real users too.;-)
Nikita recently invented and implemented a clever bit of code that keeps track of the highest node in the tree that spans a directory, and then performs repeat lookups within the same directory starting from there rather than the root. This is a nice answer to those who keep asking me, wouldn't it be faster to have separate trees for each directory? Now I have better answer for them --- nice work Nikita. It also has the nice side effect of reducing spin lock contention on the root node for 4-way SMP.
I am hoping to move my laptop to SuSE 9.0 running reiser4 sometime this week, and I am hoping we will ask for more outside testers to help us find bugs at that time. While I have mentioned only the performance flaws in this email, our overall performance seems to leave little doubt that the filesystem as-is is far better than V3, and even though it will get much faster with another year or so of tuning, if now we are the fastest available on Linux, we should be shipping now (assuming we find no new bugs in the last round of internal testing).
Benchmarks can be found at www.namesys.com/benchmarks.html
As you can see in those benchmarks, in V4 tails IMPROVE performance due to saving IO transfer time. This is a great improvement over V3, and generally speaking V4 stomps all over V3 performance. It also scales better, has plugins, and improves semantics a little bit (big semantic improvements will be in the next major release not V4).
You'll also notice that we increased the size of the fileset to be more fair to ext3, and we tested some ext3 configurations Andrew Morton suggested testing.
--
Hans