What ever happened to Tandem and NonStop OS 2015
 |
Greg Menke
OTOH, we had a situation where the floating point context switch wasn't complete due to a poorly documented fpu register, leading to a case where a particular sequence of task switches and...
Oddly enough the compiler weenies probably get there first too... What do you think happens when a decent compiler + OS combo trips over a FP error ? The decent combo has the compiler do the trap & handle. Some languages require that facility, why do a context switch if you don't have to, eh ? ;)
You have to care in a variety of ways if you intend to operate a internetworking protocol, or even a protocol that...
Compiler weenies have to do that across a far wider range of situations. Caches are critical to performance these days, compiler writers go to considerable effort to ensure *arbitary* source code doesn't trip over the cache. OS weenies can significantly reduce the number of cases they need to handle. Compiler writers are expected to do stuff like predication and cache pre-fetching these days ... OS weenies just lock the cache lines they want for a specific *known in advance* chunk of code.
When Compiler Weenies are targetting shared memory boxes life is even more interesting for them. Remember : OS Weenies handle a small number of code paths, compiler writers have to handle every conceivable code-path.
There is one area that compiler weenies don't have as much direct involvement in... TLB Debt Management - which is none trivial and hard to get right. That said there is substantial pressure and some moves towards putting that on the compiler weenie's shoulders as well.
High performance machines like BlueGene-L are 200% compiler weenie about about 0.0000001% OS weenie (because most of the nodes run a trivial single-tasking executive).