El 20/06/2013 7:44, Adrian Yu escribió:

I'm training multiple networks based on a single database.

So to accelerate speed and reduce disk reading, I use shared_memory_object class provided by boost. Since the lab workstation is currently unavailable, I migrated my code to my personal computer.

On the lab workstation, the host program successfully reads all data to memory. But on my PC, strangely it creates a file on system drive rather than storing the data in memory. The whole database is about 3.7 GB; the lab workstation has 32 GB memory and runs Windows Server 2008 R2; my PC has 8 GB memory and runs Windows 7.

There should be enough memory to store the data. So why? Are there certain ways to force the program to keep all data in memory?

In windows, Interprocess emulates POSIX-like shared memory semantics using a file (in the server and in your workstation) instead of the native windows shared memory. However, creating this file is not very different from the native shared memory since Windows must allocate enough space in the paging file to store the whole segment in case other memory requests require dumping the shared memory to the paging file to free phyisical memory for other uses. So, you always need disk-based backup space for shared memory. If there is enough free memory in the system the OS will try to avoid writing to file, just like with native shared memory. Best, Ion