In-memory databases do not have as many latency and contention issues as disk-resident databases because they do not use disk I/O. However, in heavy workload situations, in-memory databases may suffer from other spinlock contention and in-memory access issues:
Cache spinlock contention – may become a bottleneck for in-memory caches under a heavy workload. Consider increasing the number of cache partitions to 64 or more. The additional memory resources required for large number of cache partitions is insignificant and provide improved performance by reducing the cache manager spinlock contention.
Object manager spinlock contention – if your application frequently accesses a small number of objects, you may observe spinlock contention for metadata structures, which are reported by sp_sysmon.
Bind descriptors for frequently accessed objects using dbcc tune 'des_bind' so they are never scavenged. Binding the descriptors for even a few commonly used objects may greatly reduce the overall metadata spinlock contention, and improve performance.
