Tiejun Gao
Australian National University

Using Managed Runtime Systems to Tolerate Holes in Wearable Memories

Abstract:

Emerging main memory technologies such as phase-change memory (PCM) experience
permanent failures due to wear far more quickly than DRAM. Current mechanisms
waste 98% of memory on early failures in PCM.

We propose a cooperative system with low hardware and software complexity that
uses the memory abstraction and garbage collection of modern languages to
tolerate errors transparently with far less waste. Once hardware can no longer
correct a line in a page, rather than discarding the entire page, the hardware
communicates the failure to a failure-aware OS and runtime. The runtime
salvages usable portions of the page, moving data on lines that fail during
program execution and ensuring that subsequent allocations do not use failed
portions. The hardware helps by logically clustering lines to reduce
fragmentation. Together, these mechanisms incur no performance cost before
failures and average 5% overhead for failure levels of 10% to 25%.