Slashdot Mirror
Linux May Need a Rewrite Beyond 48 Cores
An anonymous reader writes "There is interesting new research coming out of MIT which suggests current operating systems are struggling with the addition of more cores to the CPU. It appears that the problem, which affects the available memory in a chip when multiple cores are working on the same chunks of data, is getting worse and may be hitting a peak somewhere in the neighborhood of 48 cores, when entirely new operating systems will be needed, the report says. Luckily, we aren't anywhere near 48 cores and there is some time left to come up with a new Linux (Windows?)."
It appears that the problem, that affect the available memory in a chip when multiple cores are working on the same chunks of data, is getting worse and may be hitting a peak somewhere in the neighborhood of 48 cores, when entirely new operating systems will be needed, the report says.
Seriously? You picked that over my submission?
I submitted this earlier this morning I guess my submission was lacking. But if you're interested in the original MIT article and the actual paper (PDF):
eldavojohn writes "Multicore (think tens or hundreds of cores) will come at a price for current operating systems. A team at MIT found that as they approached 48 cores their operating system slowed down. After activating more and more cores in their simulation, a sort of memory leak occurred whereby data had to remain in memory as long as a core might need it in its calculations. But the good news is that in their paper (PDF), they showed that for at least several years Linux should be able to keep up with chip enhancements in the multicore realm. To handle multiple cores, Linux keeps a counter of which cores are working on the data. As a core starts to work on a piece of data, Linux increments the number. When the core is done, Linux decrements the number. As the core count approached 48, the amount of actual work decreased and Linux spent more time managing counters. But the team found that 'Slightly rewriting the Linux code so that each core kept a local count, which was only occasionally synchronized with those of the other cores, greatly improved the system's overall performance.' The researchers caution that as the number of cores skyrockets, operating systems will have to be completely redesigned to handle managing these cores and SMP. After reviewing the paper, one researcher is confident Linux will remain viable for five to eight years without need for a major redesign."
I don't know, guess I picked a bad title or something?
Luckily we aren't anywhere near 48 cores and there is some time left to come up with a new Linux (Windows?).
Again, seriously? What does "(Windows?)" even mean? As you pass a certain number of cores, modern operating systems will need to be redesigned to handle extreme SMP. It's going to differ from OS to OS but we won't know about Windows until somebody takes the time to test it.
My work here is dung.
This is exactly why people are doing research on Barrelfish (http://www.barrelfish.org/).
I thought this as well, but after more carefully reading the article, I *think* I see what the problem is. It's not really a problem with large numbers of cores in a system, so much as a problem with large numbers of cores on a chip. Since the multicore chips share caches (level 2 cache is shared, level 1 cache isn't IIRC, but I could be wrong) it's actually cache memory where the issue lies. I've worked on single system image SGI systems with 512 cores, but those systems were actually 256 dual core chips. That works fine, and assuming well written SMP code performance scales as you'd expect with number of cores.
I don't need a million points of light, just two points of multi-mode fiber and a 10 Gig-E router.
I'm not affiliated with Supermicro in any way, but they have four 1U serverboards designed for the 12 core opterons, so that's 48 cores in a 1U server. I'm guessing that Supermicro is not the only vendor of quad opteron boards supporting the latest chips. There are most likely quite a few of these in use by real people. Anyone want to speak up?
I know from personal experience that the socket F opterons performed very poorly in an 8 way configuration compared to the previous generation (socket 940 gen). I ran multiple tests on dual core chips (885s, I think), back in 2006 or 7 where I'd get nearly double the performance in going from a quad configuration to an 8 way configuration, but with the socket F breed of chips, there was no performance boost at all, it was like the clock speed was being cut in half and all the threads took twice as long to complete. I saw this behavior again and again, and the motherboard manufacturer that I was testing the chips with told me that it was an issue with the chips themselves. I think this is the reason why 8-way opteron systems are very rare now.
I'm trying to understand the point of this article..Do we really need a new paper to say that centralized memory bandwidth is at some point a limiting problem in an SMP environment? Isn't this why we have NUMA?
If you want to go after linux internals like the BKL more power to you but that horse left the stable a long long time ago as well.
You could talk about the software problem in dealing with decentralized memory access, synchronization, scalable algorithms...etc but this is all likely something needing to be addressed in application space rather than at the kernel where this paper seems to focus.
There are no shortage of huge single system image linux systems with thousands of processor cores and not a single one of them use SMP architecture. They are all NUMA based (decentralized memory access).
They did not "rewrite the kernel" for 7. They updated the code, just like every other piece of software normally does when it moves from version to version. Rewriting the kernel implies that they tore it down and started over, which is most certainly not true. Vista/2008 is NT version 6.0, 7/2008 R2 is NT version 6.1, not a rewrite.
Said, "It's just like dice but it's got more sides And it tells me who lives and who dies"
So, they found scalability problems in some microbenchmarks. Well, some of the scalability paths cited in the paper will be fixed when Nick Piggin's VFS scalability patchset gets merged. But it's not like you need to rewrite every operative system to scale beyond 48 cores, it's just the typical scalability stuff, and the kind of scalability issues found these days are mostly corner cases (Piggin's VFS being an exception).
Memory pages are reference counted. Some of the pages are shared and the cores spend a lot of time reference counting. There is a point where the reference counting overhead dominates. It is hypothesized that this could be fixed by enhancing the OS to isolate pages to physical cores thereby removing the need to reference count; this is a fundamental change to the structure of traditional virtual memory management.
What they're saying is basically two things:
First, there's a bottleneck in the on-chip caches. When a core's working on data it needs to have it in it's cache. And if two cores are working on the same block of memory (block size being determined by cache line size), they need to keep their copies of the cache synchronized. When you get a lot of cores working on the same block of memory, the overhead of keeping the caches in sync starts to exceed the performance gains from the additional cores. That's not new, we've known that in multi-threaded programming for decades: when you've got a lot of threads dependent on the same data items, the locking overhead's going to be the killer. And we've known the solution for just as long: code to avoid lock contention. The easiest is to make it so you don't have multiple threads (cores) working on the same (non-read-only) memory at the same time, that just requires some thinking on the part of the developers.
Second, you only gain from additional cores if there's workload to spread to them usefully. If you've got 8 threads of execution actually running at any given time, you won't gain from having more than 8 cores. And on modern computers often we don't have more than a few threads actually using CPU time at any given moment. The rest are waiting on something and don't need the CPU and, as long as we aren't thrashing execution contexts too badly, they can be ignore from a performance standpoint. To take advantage of truly large numbers of cores, we need to change the applications themselves to parallelize things more. But often applications aren't inherently multi-threaded. Games, yes. Computation, yes. But your average word processor or spreadsheet? It's 99% waiting on the human at the keyboard. You can do a few things in the background, file auto-save and such, but not enough to take advantage of a large number of cores. The things that really take advantage of lots of cores are things like Web servers where you can assign each request to it's own core. And no, browsers don't benefit the same way. On the client side there are so (relatively) few requests and network I/O's so slow relative to CPU speed that you can handle dozens of requests on a single core and still have cycles free assuming you use an efficient I/O model. But it all boils down to the developers actually thinking about parallel programming, and I've noticed a lot of courses of study these days don't go into the brain-bending skull-sweat details of juggling large numbers of threads in parallel.
What they are talking about really reduces to a variant of Ahmdals Law, but simply put scaling is always non-linear. There will be overheads per core for communication (why is why SMP over 16 CPUs is such a headache) and overheads per core within the OS for housekeeping (knowing what core a specific thread is running on, whether it is bound to that core, etc, and trying to schedule all threads to make best use of the cores available).
The more cores you have, the more state information is needed for a thread and the more possible permutations the scheduler must consider in order to be efficient. Which, in turn, means the scheduler is going to be bulkier.
(Scheduling is a variant of the box-packing problem, which is an NP-Complete problem, but it has the added catch that you only get a very short time to pack the threads in and scheduling policies - such as realtime and core-binding - must also be satisfied in addition to packing all the threads in.)
The more of this extra data you need, the slower task-switching becomes and the more of the cache you are hogging with stuff not actually tied to whatever the threads are actually doing. At some point, the degradation in performance will exactly equal the increase in performance for the extra cores. The claim is that this happens at 48 cores for modern OS'. This is plausible but it is unclear if it is an actual problem. Those same OS' are used on supercomputers of 64+ cores, by segregating the activities in each node. MOSIX, Kerrighd and other such mechanisms have allowed Linux kernels to migrate tasks from one node to another transparently. (ie: You don't know or care where the code runs, the I/O doesn't change at all.) The only reason Linux doesn't have clustering as standard is that Linus is waiting for cluster developers to produce a standard mechanism for process migration that also fits within the architectural standards already in use.
If you clustered a couple of hundred nodes, each with 48 cores, you're looking at having around 2000+ on the system. It wouldn't take a "rewrite" per-se, merely a few hooks and a standard protocol. To support a single physical node with more than 48 cores, you might need to split it into virtual nodes with 48 or fewer cores in each, but Linux already has support for virtualization so that's no big deal either.
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
An E10K is a glorified network computing cluster.
It's not what's being discussed at all.
A Pirate and a Puritan look the same on a balance sheet.
Yes, you can play smooth full-screen video in Linux with the "Square" preview release (which includes 64 bit support). Full-screen 720p video only uses 30-40% of the CPU on my crappy Intel graphics chip, and it's completely smooth.
The K42 project at IBM Research investigated the benefit of a complete OS rewrite with scalability to very large SMP systems in mind. This is an open source operating system supporting Linux-compatible API and ABI.
Their target systems, "next generation SMP systems", back in 2003 seems to have become the current generation of SMP/multi-core systems in the meantime.
The problem isn't scaling to that number of cores but the overhead in doing so. That's what i took from it
http://soylentnews.org/~tibman
I think specifically they are talking about having 48 cores behind an L2 cache. Or 48 cores on a single die. Multi-CPU boxes generally communicate between CPU dies via the bus and from what little I can gather, that helps reduce or eliminate the issue they're describing..
OpenSolaris is dead. Solaris sucks to use without GNU userland anyway and being sued by oracle is no fun. Besides you troll, this would not need a new linux, just some small changes to the current one.
Tilera Corp. already has CPU architecture with 16-100 cores per chip.
TILE-Gx family
Support for these is already being included in the mainline kernel.
I have 34 systems which have 48 cores already in the server room. These are quad socket systems with 4 AMD 12-core CPU's. So I call BS to the guys who think we have plenty of time, because there are plenty of people deploying these things already.
We were all warned a long time ago that MS products sucked, remember the Magic 8 Ball said, "Outlook not so good"
The point is the article dealing with a simulated theoretical cpu with 48+ cores on a single die with shared l2 cache.
The changes made are incremental and I imagine will be dealt with long before this actually becomes an issue when (or if) we get cpus with that many cores on a single die.
multi socket systems are already immune to this the way it is setup, you could have an 8 socket system with each cpu having 8 cores and it would not show the problems shown in the article.
In other words, business as usual, the kernel gets optimized for hardware that actually exists or will exist in the near future. 48 core single cpus are a few years away, and the changes to accomodate them don't require anything significant so I'm sure it will be dealt with at the time.