Multikernel

Cache coherence become a problem with the increase of cores.

Using message passing instead of shared memory can solve the problem.

http://docs.google.com/present/view?id=dc3nqpnt_19hsg376hs

try to explain slide 21,

Broadcast vs Multicast:

TLB shootdown – the process of maintaining TLB consistency
by invalidating entries when pages are unmapped
– is one of the simplest operations in a multiprocessor

OS that requires global coordination. It is also
a short, but latency-critical, operation and so represents
a worst-case comparison for a multikernel.

Broadcast performs badly due to the cache-coherence protocol used
by AMD64 processors. When the master updates the line, it is invalidated
in all other caches. Each slave core then pulls the new copy from the master’s
cache.

Multicast protocol, the master sends a URPC message to the first core of each processor,

which forwards it to the other three cores in the package.
Since they all share an L3 cache, this second message is much cheaper.

IP Loopback

Table 4 shows that Barrelfish achieves higher throughput,
fewer cache misses, and lower interconnect utilization,
particularly in the reverse direction from sink to
source. This occurs because sending packets as URPC
messages avoids any shared-memory other than the
URPC channel and packet payload; conversely, Linux
causes more cache-coherence traffic for shared-memory
synchronization.
Barrelfish also benefits by avoiding kernel crossings.