4.3 Garbage Collection Overhead
In this section, we first show the performance of the
four protocols under different transaction abort ratios
and then analyze the garbage collection overhead.
Transaction Throughput. Figure 9 and Figure 10 show
the transaction throughput of all the commit protocols
under different abort ratios with strict and no-pageconflict
isolation levels, respectively. From the two figures,
we make two observations:
(1) LightTx and AW have stable performance under
different abort ratios, while SCC and BPCC have
unstable performance. This is because more aborted
transactions incur higher garbage collection overhead in
SCC/BPCC, which will be discussed in the following
section. In contrast, LightTx and MHL-D simply discard
the aborted pages, and AW and MHL-M discard the
dirty volatile mapping table on aborts. As a result, these
four protocols are not sensitive to the abort ratio.
(2) LightTx outperforms all other commit protocols
when abort ratio is non-zero and has performance comparable
to SCC/BPCC when abort ratio is zero. In Figure
9, LightTx outperforms AW and MHL-M by 1.5%
and 7.4%, respectively. The performance cost of AW and
MHL-M mainly comes from mapping persistence (writeback
of the FTL mapping table into the flash device).
MHL-D has poor performance with strict isolation. This
is because all pages in one transaction are appended
in the same block, but only one transaction is executed