[comp.arch] Info needed on NCR 32/800

nelsn@houxj.UUCP (M.NELSON) (06/12/87)

      It has been a few months since NCR announced their first microprocesser 
based multiprocessor (NCR Tower 32/800).  
Has anyone seen it yet ?  Is there any
benchmark data available yet ?
      From the sales literature I have, it is a Multibus II based
message passing multiprocessor and it does not have any shared memory.
Could anybody shed some light on how shared memory is supported in the system
(under UNIX).  How does the operating system work ?  
How does it do load balancing ? How about message passing  ? 
Is it the way to go ?
Any other comments are appreciated. 

Thanks.

-- 
Matthew Nelson
AT&T Information Systems
ihnp4!houxj!nelsn
nelsn@houxj.ATT.COM

alvitar@madhat.UUCP (Phil Harbison) (06/14/87)

In article <220@houxj.UUCP>, nelsn@houxj.UUCP (M.NELSON) writes:
> It has been a few months since NCR announced their first microprocesser 
> based multiprocessor (NCR Tower 32/800).  Has anyone seen it yet ?  Is
> there any benchmark data available yet?

See  "Inside Edge", page 21 in the June issue of UNIX/World.  The 32/800
is a multiprocessor system based on the MC68020.  That article lists the
performance  of  various  NCR systems relative to the performance of the
Mini-Tower.   A  32/800 with 4 Applications Processors (APs) has a rela-
tive  performance  of 14.4.  A single AP configuration has a performance
of 4.5. 

> From the sales literature I have, it is a Multibus II based message
> passing multiprocessor and it does not have any shared memory.  Could
> anybody shed some light on how shared memory is supported in the system
> (under UNIX).  How does the operating system work?  How does it do load
> balancing? How about message passing?

The  32/800  uses  both Multibus-I and Multibus-II.  The logical address
space for each process is 4-Gbytes, split evenly between user and kernel
space.  The kernel space is further split into 64 subspaces of 32-Mbytes
each.   One  subspace is allocated for each AP, and one space is divided
among  the  I/O processors (1-Mbyte each); therefore, any AP can address
the space of any other AP or I/O processor.

Dynamic  load  balancing is implemented by assigning each new process to
the  least-loaded  AP.   A  static  load balancing daemon reads the load
status  of  all APs and may redistribute processes.  When a process must
be  moved to another AP, the new AP copies only the minimum user structs
and relies on normal paging to do the rest. 

The AP has a 16.7MHz 68020, supported by a 68881 FPU, an MMU, an 8-Kbyte
instruction  cache  and  a 2-Kbyte data cache.  The I cache uses virtual
addresses,  while the D-cache uses physical addresses.  The MMU is based
on  the  68861,  a gate-array that uses an off-chip translation cache to
implement a subset of the 68851 PMMU features.  The translation cache is
split between 512 kernel entries and 512 user entries, implemented in 30
nanosecond  static RAM.  Each AP can have up to 16-Mbytes of memory on a
private bus. 

Each  processor other than the AP has 1-Mbyte of private RAM.  There are
Termincal  Processors  (TP),  File  Processors (FP), Communications Pro-
cessors (CP), and a LAN Processor (LP).  The FP contains two 68010s, one
to  handle  file  system code and one to handle physical I/O.  The TP is
similar with one 68010 handling the eight ports and one executing kernel
code.   The  CP  supports  two synchronous ports, with a 68010 executing
driver  and  protocol  code,  and an 8085 handing the device interrupts.
The  LP  has  a  68010  executing  kernel code, while an 80186 and 82586
handle the Ethernet.

Enough plagiarism! :-)  For more info, consult UNIX/World.

-- 
Live: Phil Harbison
USPS: 3409 Grassfort Drive, Huntsville, AL 35805-5421
Uucp: {clyde,uunet}!madhat!alvitar
Bell: 205-881-4317

wingard@ncrcae.Columbia.NCR.COM (Steve Wingard) (06/15/87)

In article <219@madhat.UUCP> alvitar@madhat.UUCP (Phil Harbison) writes:
>
>The  32/800  uses  both Multibus-I and Multibus-II.  The logical address
>space for each process is 4-Gbytes, split evenly between user and kernel
>space.  The kernel space is further split into 64 subspaces of 32-Mbytes
>each.   One  subspace is allocated for each AP, and one space is divided
>among  the  I/O processors (1-Mbyte each); therefore, any AP can address
>the space of any other AP or I/O processor.
>

Just to clarify a point -- the Tower 32/800 does NOT use Multibus I.
Processor modules have two interfaces on them:  the Multibus II interface
which connects the processor to the rest of the system and a proprietary
interface to the processor's local memory.


Steve Wingard                        ...!ucbvax!sdcsvax!ncr-sd!ncrcae!wingard
NCR Corp., E&M-Columbia              ...!decvax!mcnc!ncsu!ncrcae!wingard
					wingard@ncrcae.NCR.COM