RFR(s): 8150362: G1 base elapsed time prediction is wrong because rs_length prediction is wrong

sangheon sangheon.kim at oracle.com
Tue Mar 8 18:08:10 UTC 2016


Hi all,

Could I have some reviews for this change?

We predict remembered set length as follows from 
G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards):
size_t rs_length = predict_rs_length_diff();

This is incomplete and it is inconsistent when compare with the use case 
of the function from 
G1CollectorPolicy::calculate_young_list_target_length().
'adj_rs_lengths = rs_lengths + predict_rs_length_diff()'
In this case, we use rs_length + rs_length_diff.

I am proposing to use the sum of rs_length and rs_length_diff.
i.e. 'rs length + rs length diff'.

CR: https://bugs.openjdk.java.net/browse/JDK-8150362
Webrev: http://cr.openjdk.java.net/~sangheki/8150362/webrev.00
Testing: JPRT, some performance tests showed no big difference.

Thanks,
Sangheon


More information about the hotspot-gc-dev mailing list