LightTx and MHL-D is less than 0.194 seconds for all
zone size settings, which is close to the recovery time in
AW/MHL-M, 0.126 seconds. The recovery time has two
components: page scanning and state checking. As processing
in the CPU (required by state checking) is much
faster than flash accesses (required by page scanning),
the page scanning time dominates the recovery time.
AW/MHL-M have the smallest recovery time because
the mappings are persistent for each transaction and
only the mapping table needs to be read at recovery.
Comparatively, SCC and BPCC require a whole device
scan because all the pages of the latest version should be
found to determine the transaction states [4]. 7 However,
the need to scan the whole device leads to a large overhead
(not to mention it increases linearly with device
capacity). LightTx tracks the pages whose mappings
have not been persistent in the Available and Unavailable
Zones, so that only pages in the two zones need to be
read for recovery in addition to the persistent mapping
table read. As a result, the recovery time overhead of
LightTx/MHL-D depends on the number of pages in the
Available and Unavailable Zones, which is much smaller
than that in the entire device, leading to significantly
lower recovery time than SCC and BPCC.
Mapping Persistence Overhead. We measure the mapping
persistence overhead using the metric of mapping
persistence write ratio, which is the number of
writes to ensure mapping persistence divided by the
total number of writes in the trace. Figure 19 shows
that mapping persistence overhead in AW/MHL-M is
3.70%, while that in LightTx is less than 0.75% for
all zone size settings. SCC and BPCC trade recovery
time for mapping persistence overhead and have no
persistent mapping table, which leads to zero overhead
for mapping persistence. AW/MHL-M uses the other
extreme approach, which makes mappings persistent for
each transaction, leading to the highest overhead. As
stated above, LightTx relaxes the mapping persistence
of pages by tracking them in zones and requires writes
to maintain persistence only during zone sliding that
happens at the end of a checkpoint (as described in
Section 3.3), leading to low overhead.
Based on these evaluations, we conclude that LightTx
achieves fast recovery with low mapping persistence
overhead.
4.6 Evaluation Summary
Table 3 shows the evaluation summary ofMHL-D,MHLM
and LightTx compared with existing protocols (AW,
SCC and BPCC). While MHL-D and MHL-M support
better transaction concurrency, they are still bounded
by the number of log heads. Also, MHL-D has poor
performance because the pages in one transaction need
to be consecutively written to the same block, which