Base-level reasoning Polus Framework Reasoning

Problem determination

Input: The current state and the goals being violated.
Output: The components whose behavior needs to change and the type of change. This is expressed as a query of the form:

$$\begin{split} & \phi\;=\; \{(c, b, change)\vert\:c \epsilon ... ...lon Behavior,\: b_{new}\:=(1+ change)b_{current}\} \end{split}$$

Approach: Problem determination has been an area of ongoing research. We briefly describe a simplified approach to illustrate how problem determination, base-reasoning and meta-reasoning work together.

• Determine the components being used in the invocation path.
• Compare the current behavior of these components with the static capabilities of the components. This is similar to system diagnosis using model-based reasoning [22].
• Additionally, compare the current state with a previous state (in which the goals were met). This change analysis generates additional parameters used to search the specifications and tune components in the invocation path that are not saturated. The changes in the system state can be along one or more of the following dimensions: a) Resources utilization (accounts for failures, resource-additions, and application request rate), b) Workload characteristics c) Assigned goals, d) Invocation path (accounts for changes in active datasets or physical components).

In the SAN file system example, by analyzing the invocation path (i.e. client machine, controller, disks), we determine that disks are saturated (i.e. the current I/O rate is the maximum they can support). Furthermore, a change analysis with a previous state reveals that client I/O request rate has increased by 40% and that the sequential/random ratio of the workload has changed from $0.1$to $0.7$. Based on the problem determination analysis the following two queries get generated: Query 1: Select an action that improves the throughput of the disks by 25% (the fact that it is saturated will show-up in the preconditions of actions). Query 2: Select an action that improves the throughput of the (controller or client machine) by 25%, and is optimized for sequential workloads.

Base-level reasoning Polus Framework Reasoning