bob@wurlitzer.org wrote:
> Some motherboards have a ram qualification where they will accept a
> smaller amount of higher speed memory or a larger amount of slower.
> Would someone please give me a general idea of where the advantages of
> either approach would show up in actual use? Let's say for example
> I'm debating whether to use 4GB of DDR2 800 vs. 8GB of DDR2 533. (My
> question is assuming a functioning 64 bit operating system which can
> actually utilize more than the 4GB to which most of us are limited.)
> Thanks very much! Bob
First of all, memory has to be used, to give an advantage. If
the memory is just sitting there, and not doing anything, it
is kinda a waste.
Some operating systems, make background uses of memory. For
example, I had one computer here, with 2GB of memory in it.
One day, I ran a checksum command on a file (you see the disk
light operating while the checksum runs). I tried the command
a second time. The disk didn't light up at all. The entire
1GB file had been cached in system memory, and when I asked to
use the file a second time, the cached copy was used. So
for that operating system, the extra memory was not wasted.
The second execution of the checksum command, finished in a
fraction of the time of the first execution.
To give another example, hardware designers do logic simulation
(or event simulation). Some of those use a lot of memory.
Say, for example, I do a logic simulation of an entire silicon
chip of some sort. Say it takes 5GB or memory to hold it. One
of the characteristics of logic simulation, is locations in memory
are accessed pretty well at random, without much of a pattern.
The levels of cache on the CPU are pretty well useless for this,
because cache lines are constantly being evicted to make room
for new, randomly accessed information. Such a random access
pattern is called a "cache buster", because the cache is useless.
If I have 8GB of memory, and a 5GB problem to handle, the limiting
factor is the RAM speed. I suppose I could manage 1GB/sec of random
access without too much trouble, in that scenario. (Memory transactions
frequently happen in bursts, in cache line sizes, but it is possible
that portions of the read information are not used, so the net
memory bandwidth can be quite low. In other words, the results
you got in SiSoftware bandwidth test, won't be seen in the logic
simulation.)
Now, take the 4GB physical RAM machine. I set the swap on the machine,
so that if physical memory is exceeded, parts of the memory are swapped
out to disk. That is how 5GB of problem size, fits in 4GB of physical
memory, and 1GB of disk swap storage.
Now, what happens with a program with random access patterns. Well,
the disk swaps like crazy. In fact, I could very well find the
execution of my simulation is dominated by the speed of the disk.
Instead of the juicy 1GB/sec random access speed I had in the one
case, with the thrashing going on in the second case, I'm seeing
60MB/sec or so. That means my simulation is running 16 times
slower, due to insufficient RAM.
Not all computing problems will show such a large effect. If the
memory references have good locality, or if portions of memory
are not referenced that often, swapping to disk may not actually
be that bad.
With the small problems I work on, I want your "4GB DDR2-800" setup.
If you daily go over the 4GB resident mark, then you'd want the
"8GB DDR2-533" case instead.
Another thing to remember, is that programs have limits as to
how much memory they can access. They have to be compiled correctly,
for large or extra-large memory usage. You might waste the money
on 8GB of memory, to find that only 2GB or 3GB can be used by
any one single program. (No problem, if you wanted to load four
different 2GB programs.) You should do some research in that area,
to determine just exactly how the 8GB will get used.
Search terms - WOW64 (windows on windows emulation for 32 bit programs
on a 64 bit OS), LARGEADDRESSAWARE compiler flag. Limits can
exist at the 2GB, 3GB, and 4GB levels. (Linux will have its own
set of issues - those were for Windows.)
Paul
