when the abort ratio is high.
Sensitivity to Garbage Collection. To further understand
the garbage collection cost in LightTx versus
SCC/BPCC, we measure transaction throughput and
garbage collection time of both LightTx and SCC/BPCC
under varied garbage collection thresholds6. Figures 13
and 14 show transaction throughput and average GC
time for LightTx and SCC/BPCC with varied GC thresholds
when abort ratio is zero. From the two figures,
LightTx has transaction throughput close to that of
SCC/BPCC, and the garbage collection time is almost
as same as that of SCC/BPCC. Figures 15 and 16 show
the case when abort ratio is 10%. In this case, LightTx
has better transaction throughput and lower garbage
collection time than SCC/BPCC. When GC threshold
increases, GC time in SCC/BPCC becomes worse. GC
time in LightTx increases very slowly. This is because
the restrictions on GC in SCC/BPCC causes inefficiency
in garbage collection, either by moving valid pages due
to forced erase in SCC or by unnecessary keeping of
invalid pages for straddle dependencies in BPCC. The
inefficiency causes the free pages to be consumed much
faster, and thus leads to poorer GC performance when
GC threshold is high. Comparatively, LightTx has few
restrictions on GC and is therefore not affected by GC
threshold much.
We conclude that LightTx has lower garbage collection
overhead than other approaches mainly because 1) it
avoids frequent mapping persistence by tracking recent
updates in the Available Zone, 2) it also avoids extra
constraints on garbage collection caused by pointer-