David Holland said:I was doing a simple test of the speed of a "maths" operation and when I
tested it I found that removing the loop that initialises the data array
for the operation caused the whole program to spend twice the time to
complete. If the loop is included it takes about 7.48 seconds to
complete, but when removed it takes about 11.48 seconds.
Does anybody have a suggestion as to why this is so and whether I can
trust the results of the code as it is below?
int total = 0;
int count = 65500;
signed int data[count];
It is because of the virtual memory system. When you initialize the
array beforehand, it forces the virtual memory system to materialize
memory for the pages appearing in the array. When you do not, this
materialization happens in the timed loop, and then of course it's
You should be able to confirm this by touching some but not all of the
pages beforehand and observing the resulting timings. Hint: page size
for i386 family processors is 4096 bytes. And you won't see it at all
using DJGPP, which is DOS-based and doesn't have a virtual memory
FYI, DJGPP does use a virtual memory system by default. The primary DPMI
host for DJGPP uses hardware paging and has virtual memory support: CWSDPMI.
Also, when an DJGPP application is running in a Windows dosbox, it uses the
DPMI host and virtual memory of the OS. You can, of course, use other
non-paging DPMI hosts such as PMODEDJ (aka, PMODETSR). OpenWatcom, on the
other hand, supports many DPMI hosts and/or DOS extenders. The ones I'm
aware of are non-paging and have no virtual memory support.