diff options
Diffstat (limited to 'docs/tracing.txt')
| -rw-r--r-- | docs/tracing.txt | 46 |
1 files changed, 46 insertions, 0 deletions
diff --git a/docs/tracing.txt b/docs/tracing.txt new file mode 100644 index 0000000..a7657ff --- /dev/null +++ b/docs/tracing.txt @@ -0,0 +1,46 @@ +How And Why To Use TDB Tracing +============================== + +You can trace all TDB operations, using TDB_TRACE. It is not complete +(error conditions which expect to the logged will not always be traced +correctly, so you should set up a logging function too), but is designed +to collect benchmark-style traces to allow us to optimize TDB. + +Note: tracing is not efficient, and the trace files are huge: a +traverse of the database is particularly large! But they compress very +well with rzip (http://rzip.samba.org) + +How to gather trace files: +-------------------------- +1) Uncomment /* #define TDB_TRACE 1 */ in tdb_private.h. +2) Rebuild TDB, and everything that uses it. +3) Run something. + +Your trace files will be called <tdbname>.trace.<pid>. These files +will not be overwritten: if the same process reopens the same TDB, an +error will be logged and tracing will be disabled. + +How to replay trace files: +-------------------------- +1) For benchmarking, remember to rebuild tdb with #define TDB_TRACE commented + out again! +2) Grab the latest "replace_trace.c" from CCAN's tdb module (tools/ dir): + http://ccan.ozlabs.org/tarballs/tdb.tar.bz2 +3) Compile up replay_trace, munging as necessary. +4) Run replay_trace <scratch-tdb-name> <tracefiles>... + +If given more than one trace file (presumably from the same tdb) +replay_trace will try to figure out the dependencies between the operations +and fire off a child to run each trace. Occasionally it gets stuck, in +which case it will add another dependency and retry. Eventually it will +give a speed value. + +replay_trace can intuit the existence of previous data in the tdb (ie. +activity prior to the trace(s) supplied) and will prepopulate as +necessary. + +You can run --quiet for straight benchmark results, and -n to run multiple +times (this saves time, since it need only calculate dependencies once). + +Good luck! +Rusty Russell <rusty@rustcorp.com.au> |