Monitoring Cube Performance and Usage - Expert Cube Development with SSAS Multidimensional Models - page 332

Memory differences between 32 bit and 64 bit
Having talked about memory, it is important to make a distinction between the
32-bit and 64-bit versions of Analysis Services. Nowadays, any serious Analysis
Services deployment should use the 64-bit version. However, we might still
encounter legacy servers, or simply small deployments running on virtual machines
that use a 32-bit operating system. For all these reasons, it makes sense to discuss
the differences between the 32-bit and 64-bit versions of Analysis Services regarding
memory management.
Regardless of the physical RAM installed on a system, if Analysis Services is
running on a 32-bit operating system, it can only address a maximum of 2 GB
or 3 GB of virtual memory (it depends on whether we set the /3GB switch in the
boot.ini file; see http://support.microsoft.com/kb/833721 for further details).
This limit can be an issue for scalability and performance. First of all, even if there
was more physical memory available, a process cannot use more virtual memory
than is available in its virtual address space. Unlike SQL Server, the Analysis
Services memory manager cannot use services such as AWE to access more memory.
So as we can see, it is important that the memory manager doesn't try to allocate
more virtual memory than can be addressed because this will cause errors in
Analysis Services.
With the 64-bit version of Analysis Services, there are no practical limits for virtual
memory allocation. However, this doesn't prevent memory allocation errors, because
we can still reach the limit of virtual memory that can be allocated, which is the sum
of the total physical memory and paging file. This condition is much more frequent
on a 64-bit server than on a 32-bit one, but if we understand how this can happen, we
can also apply the right countermeasures.
Controlling the Analysis Services Memory Manager
The default settings of the Analysis Services Memory Manager work well in most
cases. However, there might be scenarios where more tuning is required to optimize
performance or simply to prevent low memory conditions.