What are the trade-offs between snooping and directory-based protocols for cache coherence?

1 Snooping protocol

A snooping protocol relies on a shared bus that connects all the caches and the main memory. Whenever a processor writes to its cache, it broadcasts the address of the modified block to the bus. Other processors that have a copy of the same block in their caches can either invalidate or update it, depending on the protocol variant. A snooping protocol is simple and fast, but it has some drawbacks. First, the bus can become a bottleneck as the number of processors and cache accesses increase. Second, the protocol requires all the caches to monitor the bus constantly, which consumes power and bandwidth. Third, the protocol is not suitable for distributed systems, where the bus is replaced by a network.

2 Directory-based protocol

A directory-based protocol uses a centralized or distributed directory that keeps track of the status and location of each cache block. Whenever a processor wants to read or write a cache block, it sends a request to the directory. The directory then grants or denies the request, and coordinates the communication between the requesting processor and the other processors that have a copy of the block. A directory-based protocol can overcome some of the limitations of a snooping protocol. First, it reduces the traffic on the bus or network, as only the relevant processors are involved in each transaction. Second, it allows for more scalability, as the directory can be distributed across multiple nodes. Third, it supports more flexibility, as different policies can be implemented for different regions of memory.

3 Trade-offs analysis

When choosing the best protocol for cache coherence, several factors must be considered, such as the size and topology of the system, the access patterns and frequency of the processors, and the design goals and constraints of the hardware architects. Generally speaking, snooping protocols can offer lower latency and higher throughput for small and uniform systems with a high cache hit rate. However, they can suffer from scalability issues as the bus becomes overloaded and coherence overhead increases with more processors and cache blocks. Directory-based protocols are more complex but can offer better performance for large and heterogeneous systems, with a network that is optimized and an efficient directory. They can also scale better, as the directory can be partitioned and distributed while reducing coherence traffic. Snooping protocols are simpler to implement and verify, while directory-based protocols require a directory structure and a network protocol.