On 27/02/2012 17:34, Johan Tibell wrote:

Hi Simon,

On Mon, Feb 27, 2012 at 12:25 AM, Simon Marlow wrote:

...

"Think of -Hsize as a variable -A option. It says: I want to use at least size bytes, so use whatever is left over to increase the -A value."

Doesn't that describe exactly what it means?

Maybe. Let me start with the mental model I approach this with: the allocation area (i.e. the nursery) should have a size in the order of megabytes, often around the size of the L2 cache.

Ah, so I see where your confusion arises - this assumption is not true in general. Just discard the assumption, and I think everything will make more sense. Picking a size for -A around the L2 cache is often a good idea, but not always. GHC defaults to -A512K, but programs that benefit from much larger sizes are quite common. For more about the tradeoff, see my SO answer here: http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll...

Given this model, I read the above as:

* if you set e.g. -H1G, you'll get an allocation area which is in the order of 1Gb large. That makes no sense to me.

Right - see above. In fact there's no problem with a 1GB nursery.

* The suggested size of the total heap (-H) has something to do with the size of the allocation area (-A). This makes no sense to me either.

So either I do understand what -H does, but it makes no sense to me, or I don't understand what -H does, but what it does makes sense.

Perhaps the confusion lies in the phrase "left over." Left over from what?

Left over after the memory required by the non-nursery parts of the heap has been deducted. Cheers, Simon

On 28/02/2012 15:59, Johan Tibell wrote:

On Tue, Feb 28, 2012 at 12:57 AM, Simon Marlow wrote:

...

Ah, so I see where your confusion arises - this assumption is not true in general. Just discard the assumption, and I think everything will make more sense.

Picking a size for -A around the L2 cache is often a good idea, but not always. GHC defaults to -A512K, but programs that benefit from much larger sizes are quite common. For more about the tradeoff, see my SO answer here:

http://stackoverflow.com/questions/3171922/ghcs-rts-options-for-garbage-coll...

Thanks for the explanation.

One has to be very careful in selecting the size of the allocation area in benchmarks. If the allocation area is large enough the GC might not need to run at all for the duration of the benchmark, while in a real program it would run.

It is a problem, yes. You also have to be careful when comparing two benchmarks runs that one didn't do an extra GC, because that can skew the results against it. I'm fairly sure the GC community have looked into this problem, but I don't know of any references off hand. Trawling Richard Jones' GC bibliography might turn up something: http://www.cs.kent.ac.uk/people/staff/rej/gcbib/ Cheers, Simon