Re: 2048 CPUs [was: Re: New filesystem for Linux]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





On Sun, 5 Nov 2006, Albert Cahalan wrote:

On 11/4/06, Mikulas Patocka <[email protected]> wrote:
>> Does one Linux kernel run on system with 1024 cpus? I guess it
>> must fry spinlocks... (or even lockup due to spinlock livelocks)
>
> The SGI Altix can have 2048 CPUs.

And does it run one image of Linux? Or more images each on few cpus?

Yes, of course it runs one image of Linux. It's SMP, not a cluster.
If I remember right:

2 CPUs per chip
2 chips per board (with local RAM; this is NUMA)
512 boards per machine

How do they solve problem with spinlock livelocks?

If time-spent-outside-spinlock/time-spent-in-spinlock < number-of-cpus,
the spinlock livelock may happen --- this condition is not true normally
with 2 or 4 cpus, but for that high amount of cpus, there is a danger.

Or do they have some special hardware spinlock instruction that guarantees
fairness?

Well first of all it's using IA-64 processors. These do atomic operations
via instructions that are essentially "take lock" and "release lock".
The non-CPU hardware probably recognizes these operations by the
special bus cycles and does whatever is needed to ensure fairness.

It looks like a normal looping spinlock with rep nop and cmpxchg on i386. Not that I understand IA64 assembler --- but look at ia64_spinlock_contention --- it looks like a loop with pause and nonatomic test and a branch to acquire spinlock atomically with cmpxchg --- nothing that would prevent starving there.

Second of all, I think we just try to avoid having long spinlock hold
times. The SGI people love using RCU. As I recall, SGI also caused
Linux to get a sequence lock for jiffies.

That is right, that you should avoid contention --- but the kernel should be stable no matter what userspace programs you run. Or does it mean that scientists can run only trusted programs on Altix that do not perform too many certain syscalls concurently in order to not jam spinlocks?

Mikulas
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

[Index of Archives]     [Kernel Newbies]     [Netfilter]     [Bugtraq]     [Photo]     [Stuff]     [Gimp]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Video 4 Linux]     [Linux for the blind]     [Linux Resources]
  Powered by Linux