As with TPC-C, the TPC-D benchmark specifies a method for scaling the database.
These auxiliary tables contain partially aggregated data which can be precomputed in the database load phase so that very little processing needs to be done when executing the TPC-D queries.
For instance, for TPC-D, each of the 17 queries is implemented as a database transaction, while the two update streams are broken down into smaller database transactions.
TPC-C TPC-D Source TPC TPC -- benchmark benchmark C D Platfrom MVS on WinNT on WinNT on IBM S/370 intel X86 Intel X86 DBMS DB2/MVS DB2/UDB DB2/UDB Date Collected -- 2/10/1998 3/8/1998 Duration (h:m) 4:15 (withheld) (withheld) # Objects 1224 101 192 Data Size (MB) 42792 70246 77824 Footprint (MB) 2957 13267 51580 # References 18550114 196067649 218130354 # Xacts 119235 890885 230 Read Ratio (%) 90.9 87.4 97.8 The production traces were taken off the primary systems in use at some of the world's largest corporations in the early nineties.
For the TPC-D trace, we used only a quarter of the trace to warm-up our simulators, since it had already achieved a large enough footprint.
Warm-Start Point Trace Aerospace Bank ConsGds DirMktg1 DirMktg2 # References 3889504 20000000 3566923 3200940 7198063 % References 50.0 55.7 50.0 50.0 50.0 % Trace Time 42.6 65.8 49.1 45.0 40.1 Trace FinSvcs Insurance Retail TelecomA TelecomB1 # References 7832002 10324437 19323180 6536458 5765598 % References 50.0 50.0 50.0 50.0 50.0 % Trace Time 51.4 50.4 50.5 64.6 45.3 Trace TelecomB2 Utility TPC-C TPC-D # References 6878687 18826685 98033825 54532589 % References 50.0 50.0 50.0 25.0 % Trace Time 50.3 50.3 51.9 42.0 In this article, we generally present only the results for buffer pools that are filled at the predetermined warm-start point.
On the other hand, the reference patterns exhibited by the TPC-D queries are clearly sequential and are more structured and regular than those of the production workloads, perhaps because the TPC-D queries are run serially.
Notice also that the plot for TPC-D's queries contains clearly sequential patterns with two or three dark horizontal lines.
Compared to the production workloads on average, TPC-C's miss ratio improves faster with increases in buffer pool size, while the opposite is true of TPC-D. For the most part, the average miss ratio of the production workloads falls between that of TPC-C and TPC-D.
To a much lesser extent, TPC-D also has this behavior but not TPC-C.