There is an open pull request for incremental GC, and Jeff has mentioned that other GC improvements are planned after 0.3 is released.
On Tue, Apr 1, 2014 at 11:55 AM, Freddy Chua <[email protected]> wrote: > Strange, although my for loop does not create any additional memory, the > memory usage increases to 60GB after turning off GC... > > On Tuesday, April 1, 2014 11:44:06 PM UTC+8, Freddy Chua wrote: >> >> Alright, these are my timings are disabling gc >> >> before disabling gc >> each for loop takes 911.240040 >> after disabling gc >> each for loop takes 30.351131 >> >> around 30x improvement.... and if my loop run 10 times, it would have >> been a 300x improvement... >> >> I hope julia devs do consider improving the GC invocation... >> >> On Tuesday, April 1, 2014 11:28:44 PM UTC+8, Freddy Chua wrote: >>> >>> Alright, I am pretty certain that >>> >>> macro nogc(ex) >>> quote >>> try >>> gc_disable() >>> local val = $(esc(ex)) >>> finally >>> gc_enable() >>> end >>> val >>> end >>> end >>> >>> Does the trick... >>> >>> My program does an iterative "gradient descent", a kind of mathematical >>> optimisation algorithm. So I loop through a for loop multiple times. In >>> each loop, nothing gets created or destroyed, so GC is not needed at all. >>> It turns out that turning off GC improves the performance significantly, >>> probably 100x. This issue is serious, I wonder can there be a better way of >>> determining when to call GC. I guessed disabling GC manually is not the >>> intention of the compiler designers.. >>> >>> On Tuesday, April 1, 2014 11:12:07 PM UTC+8, Freddy Chua wrote: >>>> >>>> I found this, https://groups.google.com/forum/#!searchin/julia- >>>> users/garbage/julia-users/6_XvoLBzN60/EHCrT46tIQYJ >>>> >>>> Might try to turn off GC and see whether performance improves, will >>>> update here later... >>>> >>>> On Tuesday, April 1, 2014 11:07:03 PM UTC+8, Stefan Karpinski wrote: >>>>> >>>>> This code doesn't seem to create a List, Nodes or insert them into a >>>>> Dict - it just walks over a preexisting linked list. >>>>> >>>>> >>>>> On Tue, Apr 1, 2014 at 10:59 AM, Freddy Chua <[email protected]>wrote: >>>>> >>>>>> A possible flaw I have is the circular dependency in the data >>>>>> structures between Bus_Stop and Edge. >>>>>> >>>>> >>>>>
