7. Conclusions

 

We have described the centerpiece of a server architecture for designing high-performance LAN servers. The server is separated into generic cache modules and custom modules. The generic module is policy-free and implements general and optimized mechanisms to handle cache requests. The custom modules can enforce arbitrary policies on the management and use of the cache (including authentication and object consistency). For cache-friendly applications, the resulting servers can perform an order-of-magnitude faster than existing systems. Since the custom modules can implement application-specific cache management policies, the likelihood that an application is cache-friendly can be increased.

We have learned that reducing main memory conflicts between the CPU and the network conrollers is the key to achieve high network throughput. Therefore, we provide a detailed examination into how the cache module must be designed to make efficient use of all hardware components.

There remain many open questions about how to use this server architecture. It is difficult to determine the cache-friendliness of applications designed to support thousands of clients in a laboratory setting. In the future, we plan to investigate the breadth of applicability of this architecture to current applications and investigate new classes of applications that may be enabled by our architecture.