Web Service Models

  

Figure: A simple model for serving static Web content. Requests arrive at rate $\lambda$ (which varies with time) and incur an average service demand D_P in a CPU. An in-memory cache of size M absorbs a portion H of the requests as hits; the misses generate requests to storage at rate $\lambda _S$.

This section presents simplified analytical models to predict performance of Web services with static content, based on the system model depicted in Figure 3. Table 1 summarizes the model parameters and other inputs. Table 2 summarizes the model outputs, which guide the choices of the resource allocator described in Section 4.

The models are derived from basic queuing theory and recent work on performance modeling for Web services. They focus on average-case behavior and abstract away much detail. For example, they assume that the network within the data center is well-provisioned. Each of the models could be extended to improve its accuracy; what is important is the illustration they provide of the potential for model-based resource provisioning.

 

Table 1: Input parameters for Web service models.

Parameter	Meaning
$\alpha$	Zipf locality parameter
$\lambda$	Offered load in requests/s
S	Average object size
T	Total number of objects
M	Memory size for object cache
DP	Average per-request CPU demand
$\mu_S$, $\phi$	Peak storage throughput in IOPS