I think we are at a stage where we should move into some performance
fixes. Both single processor, and scalability. Here's a list that I
came up with within a few minutes. All of these items are mips
or origin specific work.
* kernel text replication (implemented as a CONFIG option, perceived
as important for scalability).
* improved tlbmiss handlers (single processor performance, will let
us compare baseline results with irix for _simple_ apps). Improved
locore handlers.
* multithreaded and improved ethernet driver for networking apps
(Ralf was looking into this?)
* per node local memory (possibly for holding the locore handlers,
like IRIX?, might help scalability)
* distribute (per node) global arrays, most probably the io
arrays (leave this for Leo), as well as the nasid-cpu-cnode
translation arrays. Look into tlb wired entries per processor.
Start having more per processor and per node information.
Devise a method to access current node information in a quick
way.
* start running _simple_ benchmarks (Ulf indicates a problem
with lmbench context switching).
Hopefully, some of this can also be picked up for the snia port.
Btw, the mips64 kernel has booted up multiuser the machine posix0.
It can see 32 cpus, access all the memory on the 16 nodes, and
understands there are 11 PCI busses with 24 scsi controllers and
13 disks.
Kanoj
|