There may not be enough virtual memory when you have a large number of databases in SQL Server

On a computer with 2 GB or more of RAM, SQL Server reserves all but 256 MB (SQL Server 7.0) or 384 MB (SQL Server 2000) of virtual address space during the startup process for use by the buffer pool. Additionally, to storing the data and procedure cache, SQL Server uses the buffer pool memory to service most other memory requests from SQL Server processes that are less than 8 KB. The remaining unreserved memory is intended for use with other allocations that cannot be serviced from the buffer pool. These allocations include, but are not limited to:

• Stacks and the associated thread environment block for any threads that SQL Server creates. After SQL Server creates all 255 worker threads, this is approximately 140 MB.
• Allocations that are made by other DLLs or processes that are running in the SQL Server address space (which varies from system to system), such as:
  
  • OLE DB providers from any linked servers.
  • COM objects that are loaded by use of the sp_OA system stored procedures or extended stored procedures.
• Any images (.exe or .dll) that are loaded in the address space, which commonly use 20 to 25 MB, but possibly more if you are using linked servers, sp_OA, or extended stored procedures.
• The process heap and any other heaps that SQL Server might create. During the startup process, this is typically 10 MB, but may be more if you are using linked servers, sp_OA, or extended stored procedures.
• Allocations from SQL Server processes that are greater than 8 KB, such as those required for large query plans, send and receive buffers if the network packet size configuration option is close to 8 KB, and so on. To see this number, look for the OS Reserved value that is reported in DBCC MEMORYSTATUS and that is reported as number of 8-KB pages. Typical values for this are 5 MB.
• An array to track status information for each buffer that is in the buffer pool. This is typically about 20 MB, unless SQL Server is running with Address Windowing Extensions (AWE) enabled, in which case it can be significantly higher.

On systems that have a large number of databases, the 64-KB allocations that are required for log formatting might occupy all of the remaining virtual memory. At that point, subsequent allocations may fail, resulting in one or more of the errors that are listed in the "Symptoms" section in this article.

By using the -g startup parameter, you can instruct SQL Server to leave additional virtual memory available so that the combination of these log-related allocations and other normal allocations does not run out of virtual address space.

The following table lists some suggested starting points for the -g value depending on the number of databases and the server version: This table was calculated by using the typical values that are listed and it is also based on the assumption that no linked server activity, sp_OA or extended stored procedures are in use. It is also based on the assumption that you are not using AWE, and that SQL Profiler is not in use. Any of these conditions may require you to increase the value of -g.

Microsoft recommends that you take serious consideration before you run a server with more databases than this because the overhead that is required for having this number of databases on the system is taking a lot of virtual memory away from the buffer pool, which may result in poor performance for the system as a whole.

Additionally, creating lots of databases has the most significant effect on virtual memory. There are also per-database memory allocations that may cause a buffer pool out-of-memory condition. For example, you may receive the following error message:This kind of out-of-memory condition may be more common in SQL Server 2005 because SQL Server 2005 tracks more per-database metadata than earlier versions of SQL Server.

When you track the index usage by using the sys.dm_db_index_usage_stats stored procedure, the operation may require a significant amount of memory. The operation requires a significant amount of memory if each database has lots of indexes.