5.1 Adding Server Modules

There are three methods for adding server modules: (a) add miss-servers either to decrease miss latency or to handle heterogeneous objects, (b) add hit-servers to increase the cache size and hit rate, and (c) add hit-servers to improve the total bandwidth and handle more clients.

(a) Scalability is aided by an explicit separation of miss-server and hit-server hardware: miss-server CPU and IO consumption does not degrade hit-server throughput. Miss handling only influences the throughput of the hit-server when the miss-server stores an object into the hit-server cache.

(b) For certain cases, the bandwidth of a single hit-server might be sufficient but its main-memory cache might be too small for the application's working set. In particular, this can happen if the hit-server's motherboard supports less memory than the processor can address. Then, multiple hit-servers can be used to increase the object cache. Each hit-server holds the entire cache directory but the cached objects are partioned among all hit-servers. Client requests are multicasted to all hit-servers. If the request hits, the according hit-server executes it; the other ones classify the request as a hit but ignore it since they do not have the object. If a request misses, all hit-servers detect a miss and a dedicated hit-server signals it to the miss-server. This one then selects a hit-server for loading the new object. Fortunately, no complicated consistency protocol is required for this type of scaling. Network load, miss-server load and hit-server bandwidth are identical to the single hit-server case; only the resulting cache size is increased.

(c) When the number of clients becomes too large, hit-servers must be scaled to increase the total bandwidth of the system. This is simple as long as all objects are read-only. As soon as objects are write-shared between multiple hit-servers, we need consistency protocols.