IBM SG24-4576-00 Server User Manual


 
Dedicated Cache:
This SMP design supports a dedicated L2 cache for each
processor. This allows more cache hits than a shared L2 cache. Adding a
second processor using a dedicated L2 cache can improve performance as
much as 80%. With current technology, adding even more processors can
further increase performance in an almost linear fashion up to the point where
the addition of more processors does not increase performance and can actually
decrease performance due to excessive overhead.
The IBM PC Server 720 implements SMP with dedicated caches.
Figure 2 shows SMP with dedicated secondary cache.
┌───────────────┐ ┌───────────────┐
│ Pentium │ │ Pentium │
└───────┬───────┘ └───────┬───────┘
││
││
││
┌────────────┴─────────────────┐ ┌──────────────┴────────────────┐
│512KB Secondary (level2) Cache│ │512KB Secondary(level2) Cache│ │
└────────────┬─────────────────┘ └──────────────┬────────────────┘
││
││
││
└────────────────────┬──────────────────┘
┌────────┴─────────┐
Main memory
└────────┬─────────┘
Figure 2. SMP with Dedicated Secondary Cache
Dedicated caches are also more complicated to manage. Care needs to be
taken to ensure that a processor needing data always gets the latest copy of that
data. If this data happens to reside in another processors cache, then the two
caches must be brought into sync with one another.
The cache controllers maintain this
coherency
by communicating with one
another using a special protocol called MESI, which stands for
M
odified,
E
xclusive,
S
hared, or
I
nvalid. These refer to tags that are maintained for each
line of cache, and indicate the state of each line.
The implementation of MESI in the IBM PC server 720 supports two sets of tags
for each cache line, which allows for faster cache operation than when only one
set of tags is provided.
1.3.2 Memory Interleaving
Another technique used to reduce effective memory access time is interleaving.
This technique greatly increases memory bandwidth when access to memory is
sequential such as in program instruction fetches.
6 NetWare Integration Guide