Snapshot-based loading acceleration of web apps with nondeterministic javascript execution

Abstract

JavaScript execution is heavily used during the loading of web apps, taking a substantial portion of the app loading time. To accelerate JavaScript execution, snapshot-based app loading has been proposed [5, 17]. We take a snapshot of the JavaScript objects in the heap at some point during app loading (which we call snapshot point) and save them in a file in advance. Then, we start app loading by copying the objects in the snapshot to the heap directly, skipping JavaScript execution to create those objects. One issue is that the JavaScript execution state at the snapshot point should be the same at every app loading. If JavaScript execution included in the snapshot is involved with some nondeterminism (e.g., use random function or current time/location), snapshot-based app loading might be inapplicable since the loaded state might differ each time. In this paper, we perform an empirical study for the nondeterministic behavior of web apps during app loading. The result shows that nondeterminism is used frequently, but its impact is small from the user's perspectives. Based on this observation, we propose two techniques that can increase the applicability of snapshot. First, we propose two types of snapshot point, which allows saving as much execution state as possible in the snapshot, just before the nondeterministic execution affects the user's perception. Second, we develop a tool to identify those snapshot points by static and dynamic analysis, so that the developer can easily identify them. Our experimental results show that the techniques can accelerate the app loading by 1.81 times, by applying the snapshot that would otherwise be inapplicable, achieving a performance comparable to that of a snapshot that completely ignores the nondeterminism issues.