Request for review (m) 8008966: NPG: Inefficient Metaspace counter functions cause large young GC regressions

Coleen Phillimore coleen.phillimore at
Wed Mar 27 14:14:34 PDT 2013

Hi Jon,

In metaspace.hpp maybe capacity_in_bytes should be called 
capacity_bytes_slow() or something like that to distinguish it from 
capacity_bytes() which is fast.

*+   // Total capacity in all Metaspaces*
     static size_t capacity_in_bytes() {
*       return capacity_in_bytes(Metaspace::ClassType) +*
*              capacity_in_bytes(Metaspace::NonClassType);*
*+  #ifdef PRODUCT*
*+      // Use capacity_bytes() in PRODUCT instead of this function.*
*+     guarantee(false, "Should not call capacity_in_bytes() in the PRODUCT");*
*+ #endif


Or maybe since capacity_in_bytes() is used for the other counters, 
change the capacity_bytes() name to allocated_capacity() or something 
like that.   Just so the two names are more different than the presence 
of "_in".


line 270 and 283 are missing an 'n' in accounting.   I like the promise 
of a cleanup.   Even with the comment, it's hard to keep these straight.

1199-1201 is the same code as above it.

2520 capacity_in_bytes(mdtype) is still called for PrintGCDetails which 
iterates over the CLD graph.   This seems too expensive for GC 
printing.   It also calls used_in_bytes() so iterates twice. Then x2 for 
class vs. data metaspace.   This wasn't part of the GC slowdown that was 

2935.  I don't understand why we are checking UseMallocOnly since we 
don't use malloc for metaspaces ever.

I can't comment on the CMS change.  It looks like you just moved it.


On 3/27/2013 12:13 AM, Jon Masamitsu wrote:
> Replace the use of a method that calculated the total capacity of
> the Metaspaces by iterating over all the Metaspaces by maintaining
> the sum of Metaspace capacities as the Metachunks are
> allocated to each Metaspace.  Also maintain a sum for each
> Metaspace as the Metachunks are allocated to that Metaspace.
> Change should_expand() and compute_new_space() to
> calculate quantities in bytes.
> Remove calls to methods that calculated totals for
> "used" and "free" by iterating over the Metaspaces
> in product mode.  In some cases substitute the use
> of capacity for used.
> Thanks.
> Jon

-------------- next part --------------
An HTML attachment was scrubbed...

More information about the hotspot-gc-dev mailing list