Hi,

nick (anon@anon.com) on 5/14/06 wrote:
>Brendan (btrotter@gmail.com) on 5/13/06 wrote:
>>nick (anon@anon.com) on 5/13/06 wrote:
>
>>>No, it would suck. You're about 10 years behind: lock
>>>contention doesn't hurt scalability any more, cacheline
>>>contention does.
>>
>>Am I?
>>
>>For my last prototype, each thread has a "thread control block" containing it's
>>message queue lock (one queue per thread). While each lock wasn't alone in it's
>>own 128 byte area (Intel's recommendation) the data around
>
>That's because it is the size of their largest dcache line.
>
>it was deliberately arranged
>>so that only rarely accessed things are in the same area as the lock. The chance
>>of 2 or more CPUs trying to read/write to any of the items in the same cache line
>>as the lock (including the lock itself) is very small.
>
>Doesn't matter, if they have to both take the lock ->
>cacheline bouncing; if one CPU writes a messages which is
>read by another -> cacheline bouncing.

That would involve a cache line fill. It's not even close to the performance effects of lock contention (which either involves repeated cache line fills or context switches and TLB flushing).

>>Given that (IIRC) Linux itself used a "big kernel lock" and nothing else for version
>>2.0, and that this lock is still used (although for a lot less than it originally
>>was), and that the slowest kernel function within Linux 2.0 would've been much longer,
>>I don't think you're in a position to argue too much.
>
>Err, I don't follow your logic at all. Linux 2.0 has a
>global lock, scalability sucked. What is my position, and
>why isn't it one to argue too much? How much am I allowed
>to argue?

Linux 2.0 used a big kernel lock that would've been acceptable for that release (given developer time constraints and a probable desire to support 2-way SMP before worrying much about larger systems). Sometimes scalability isn't the most important factor - if it was they probably wouldn't have used a big kernel lock to start with.

Still, I'm getting "side-tracked". My original comment was about the time taken by the longest kernel function. Consider a system where a CPU spends 50% of it's time in the kernel on average, compared to a system where a CPU spends 1% of it's time in the kernel on average. For the first system "big kernel lock contention" is extremely likely (even for a 2-way system), while for the second system "big kernel lock contention" is unlikely (even for 8-way SMP).

For monolithic systems the time spent in the kernel includes the time spent within all device drivers, so the time spent inside the kernel is going to be greater than an equivelent micro-kernel (where time spent in the kernel doesn't include the time spent within device drivers). Therefore using a big kernel lock is going to effect performance/scalability in a monolithic system much more than it would effect performance/scalability in an equivelent micro-kernel.

>>>>For control information you've got a choice - access the control data across nodes,
>>>>or access the device driver code and it's local state across nodes. Unless everything
>>>>(control data, code and state) is on the same node it can't be "perfect" regardless
>>>>of what you do (for all kernel designs).
>>>
>>>That isn't a choice, that is a restriction!
>>
>>It's a hardware restriction that can't be avoided, where software developers can
>>choose a way to attempt to minimize the "inter-node" access penalties. I will assume
>>that for Linux, you end up with penalties when accessing device driver code and
>>the device driver's local state (and no penalties when accessing the control data).
>>
>
>In Linux, the control data does not need to go across
>node because it is basically kept in the stack of the
>running process. Instruction text is a different matter,
>but AFAIK, that can be replicated on the big NUMA machines
>in Linux.

My last prototype used multiple copies of the kernel (one for each NUMA domain) and my current prototype uses multiple copies of the kernel and multiple copies of "static kernel data" (anything that doesn't change after boot). My kernels are normally less than 64 KB though so the memory costs aren't too bad and it's still sensible when the "across node penalty" isn't too high (e.g. 2 or 4 way Opterons). Linux is typically around 3 MB so you might get better results using this memory for disk cache instead. :-)


Cheers,

Brendan