Using Table Splitting

Question

Hello,
I am doing experiments in preparation of a unicode conversion of a 1 TB MAXDB system. Therfore I have exported the database with sapinst with a duration of 44 hours.
To reduce the downtime I did another try using the table splitting option with the export monitor. Because I have not found detailed information I did a mistake in konfiguration of the file import_monitor_cmd.properties, so the export monitor did not use table splitting but needed also 44 hours for export (in addition to calculation of table splitting of around 20 hours which can be done online).
At third try I konfigured the export monitor right, the duration increases to 7 days.
Technical details:
R3 Enterprise EXT 1.1, Kernel 640, MAXDB 7.5, actual R3ta
10 CPU, 30 GB RAM, SAN
8 parallel processes for export,
8 tables to split to 16 parts each (COEP, COSS, COSP, COSB, GLPCA, CDCLS, AUSP, IBINVALUES)
largest table: COSS with 85 GB
DB-Size filled: around 1 TB, 50 Datafiles
In special of MAXDB it would be best to use primary key for table splitting. But I found no documentation of the file R3ta_hints.txt where I can manipulate the splitting decision of R3ta. (I added IBINVALUES IN_RECNO as selective column).
And now the questions:
Do you know a detailed dokumentation of the file import_monitor_cmd.properties?
Do you know a documentation of the R3ta table splitting options and input possibilities of R3ta_hints.txt?
Can you give me some experiences of similiar tests or migrations?
Best regards
Andreas

Answer 1

I´m actually trying to do a likewise thing with a 2,1 TB MaxDB 7.6.
(in addition to calculation of table splitting of around 20 hours which can be done online).
I´d be careful doing that online. If either at the time R3ta starts to evaluate the distinct values or after that table is finished someone does an insert to one of those tables you may miss that entry.
I used DB50 - Statistics to find out the biggest tables. After I had those I created my own  R3ta_hints.txt checking each table in SE11 and DB50 to find out which field has the most distinct values. I have 85 tables in the R3ta_hints.txt.
The R3ta run itself took almost three days (though database prefetching was activated).
Since I had good experiences with packages of 250 MB in former migrations I spllited the rest of the tables with that value. At the end there were 1,114 packages.
I started the export last friday, it´s still running.
After my experiences the MaxDB can´t leverage the parallel unload as it would be expected. You may see a little better speed if you set the parameter LRU_FOR_SCAN to YES which will then use the full available data cache for table scans (as R3load does) instead of just 10 % of the cache if set to NO.
What you see if you use parallel unload is (using x_cons  sh ac 1)
T204   7     -1 User      21781 IO Wait (R)           0 0     45        98030278(s)
T206   7     -1 User      21911 IOWait(R)(041)        0 0               98030278(s)
T222   8     -1 User      21887 IOWait(R)(041)        0 0               132941587(s)
T228   8     -1 User      21775 IOWait(R)(041)        0 0               132941587(s)
T234   9     -1 User      21899 IOWait(R)(041)        0 0               105696431(s)
T236   9     -1 User      21893 IOWait(R)(041)        0 0               105696431(s)
T248  10     -1 User      21905 IO Wait (R)           0 0     40        101974717(s)
T249  10     -1 User      21311 IO Wait (R)           0 0     41        101974717(s)
The IOWaits with 041 are no real reads but in fact waits for another task reading the same tree/leafs.
Markus