GC Method (was Questions for Discussion)
Wed, 28 May 1997 07:42:15 -0700
Well, I've been reading up on GC and Baker's Treadmill with incremental-update
and write-barrier synchronization seems promising. There's a paper on this
flavor (Wilson and Johnstone 1993). Performance problems with existing
implementations may be due to the use of smart pointers, according to Jones and
Lins (Garbage Collection, Wiley, 1997). I'm thinking that if you don't use
smart pointers, this might go away.
OTOH, Nettles and O'Toole's replicating collector might also be good. FWIW,
Jones and Lins claim that Nettles and O'Toole's replicating collector (NORC)
and Baker's Treadmill (BT) are the most promising candidates for real-time GC
among software-only GC algorithms so far. There are worst-case timing for NORC
at 50 microsecs per atomic action. BT implementations surveyed don't yet match
this, but again they may be optimizable by avoiding smart pointers.
The views expressed are mine alone,
unless you agree with me.