Device-reserved Memory as an Eviction-based File Cache

In resource constrained embedded systems, memory reservation, which provides physically contiguous memory allocation to devices (e.g., video decoder and camera), decreases the memory efficiency of the system. The dedicated use of reserved memory by an owner device makes itself more under-utilized when the device is less frequently used. Previous approaches support on-demand reservation to relax the dedicated use of the reserved memory by allowing the kernel to exploit the reserved memory while the owner device is idle. The previous approaches, however, can either incur significant delay to the on-demand reservation or sacrifice the memory efficiency during the on-demand reservation. Due to the delay, an end user could wait for tens of seconds when to use the owner device-dependent functions. When the memory efficiency is compromised, the system will likely incur additional read I/Os. In this paper, we propose a scheme to use device-reserved memory as an eviction-based file cache called eCache. The eCache provides on-demand reservation by discarding the cached data in a contiguous memory region in eCache. From the nature of eviction-based placement policy, the memory efficiency is still preserved during the on-demand reservation because the cached data in eCache is always less important than those in an upper-level cache such as in-kernel page cache. In addition, the on-demand reservation only requires minimal time discarding cached data in eCache. When multiple reserved regions comprise the eCache, cost-based region selection improves the memory efficiency (or caching efficiency) in eCache when on-demand reservation occurs. We implemented eCache on the Nexus S smartphone and evaluated in Android workloads. The evaluation results show Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00. that the 21% of additional memory reduces read I/Os by a factor of from two to six as compared to those with memory reservation approach. The application launch performance is also improved by at most 16%. The on-demand reservation time is reduced to a few milliseconds.