MemBalancer

• Determine heap size for garbage collected runtime.
• Review the space time tradeoff - the bigger the heap size is the faster it runs.
• Although all heap size for a single heap is pareto optimal, for a system with multiple heaps it may not be.
• This reflect the age old insight that optimality of subcomponent does not imply optimality of the whole system.
• We introduce rule that gurantees global optimality.
• Under such rule, extra memory grow sublinearly (squareroot) w.r.t working memory, as opposed to linearly.
• Take memory away from big heap and give it to small heap, reflecting the law of marginal diminishing return.
• Gets good result in v8 (the javascript compiler of chrome) and racket.
• Currently being implemented by firefox, chromium, mmtk, and julia
• Note that you can still use MemBalancer without having multiple runtimes. A typical computer system is very rarely single-tenant. Even with a single runtime, the application still keep software cache, alongside os filecache, swap, and whatnot. Runtime taking more memory mean software caches of all form will suffer.