Re: Using a lot of Maps
On 23-11-2010 20:42, markspace wrote:
On 11/23/2010 5:07 PM, Arne Vajh?j wrote:
The Java memory model was specifically created to handle
systems without cache coherency, so it will survive fine.
Apps assuming that it will not be a problem on modern
systems will not.
Cache coherency isn't a concern. It's shared memory. Threads require it,
message passing systems generally eschew it. Message passing systems
generally use processes, not threads, to do their concurrency.
I think that system will still have shared main memory.
L1, L2 and L3 cache will not be shared.
There's also problem is simulating shared memory with a message passing
system. There's a fairly high overhead in messaging systems with passing
the message. Trying to copy a large block of memory -- or multiple large
blocks -- to simulate a Java synchronization action could really bog
performance. Do-able, perhaps, but not a great idea.
Read the article you linked to again: Intel is saying they want to
connect 1000's of processors but not use cache coherency. So if you fill
an array with 10000 bytes of output and then have to make it visible by
releasing a shared lock, that memory has to be sent to all systems that
may need it.
They will not need to send any data - they will need to send an
invalidate cache message.
One message, sent to N systems. In a shared memory design, this happens
automatically. In message passing, you have to pass a message to each
process. Each process has to wait for each message to receive it.
>
> It gets worse if the initial thread has updated memory hither and yon
> through out the system -- many messages to many processes.
I believe this it at the HW level, so no processes involved.
Arne