OraV7 on Alpha: psuedo-benchmark

Motorola Compound Semiconductor - 1 recently completed an evaluation of an Oracle V7 application on a DEC Alpha 3000/500. Enclosed is a terse summary of our findings. Please feel free to write if you would like more information. This summary will be cross posted on comp.sys.dec and comp.sys.osf.osf1.

The name of our application is "Synergy". Synergy is a small to medium sized database which consists of about 80 tables and currently stores a total of about 2 Gb of raw data.

These are not TRUE benchmarks since we were comparing a V6 database on one machine (an HP 847) against a V7 database on the other (DEC Alpha 3000/500). Still, all findings can be explained and general conclusions about both the performance of the DEC Alpha as well as differences in performance of Oracle V6 vs. V7 can be drawn.

1. Platform problems
2. couldn't load Synergy applications w/ shared libraries (linker core dumps)

        fix: upgrade from OSF/1 v1.2 to OSF/1 v1.3
             (This was a REAL problem.  DEC REALLY fixed it.)

  2. Application problems

1. Motif interface core dumps -- caused by unecesasry pointer casts ptr = (char *)(int)malloc(sizeof(ptrdef)) /* example */

        fix: change Synergy source code (Shame on us.)

  3. Oracle problems

1. V7 constraints stricter than V6

        fix: corrected UNIQUE constraints in table definitions
             (normal porting related problem)

     b. import utility core dumps

        fix: upgrade from Oracle 7.0.12.2 to Oracle 7.0.13
             (This was a REAL problem.  Oracle danced around it.)

     c. install program produces erratic results

        fix: install over and over and over and over and over ...
             (This was a REAL problem.  Oracle didn't admit to it.)

The HP 847 has 128 Mb of main memory and 6Gb on 4 Mb/S SCSI II drives. The DEC 3000/500 has 128 Mb of main memory and 8Gb on 15 Mb/S Velocitor drives.

BENCHMARK #1

This benchmark measured the time it took to load a datafile which was about 850Kb in size. The file was parsed and loaded into the database schema with a user written program (written in Pro*C), not with SQL loader.

This first set of benchmarks were taken with Oracle parameter DB_BLOCK_BUFFERS = 200 on both systems.

       HP 847    DEC 3000/500      Conclusions            Notes
       =======   ============      ===============        =============
real   120.0 s     158.5 s        HP    32% faster        total CPU time + I/O
user    14.0 s       8.3 s        DEC   68% faster        CPU time in non sys calls
sys      0.8 s       0.4 s        DEC  100% faster        CPU time in sys calls

BENCHMARK #2

• We increased the Oracle buffer cache from 200 blocks (2048 bytes/block) to 600 blocks on the Alpha (Oracle V7).

  With the increased Oracle cache the V7 DB (on Alpha) runs ~40% faster than   the V6 DB (on HP).

  Changing the cache size on the HP from 200 to 600 blocks had no noticeable   effect. The reason for the improvement in DEC performance is probably due   to the larger word size required to support a 64 bit architecture.

       HP 847    DEC 3000/500      Conclusions            Notes
       =======   ============      ===============        =============
real   120.0 s      84.0 s        DEC   43% faster        total CPU time + I/O
user    14.0 s       6.8 s        DEC  105% faster        CPU time in non sys calls
sys      0.8 s       0.4 s        DEC  100% faster        CPU time in sys calls

• Below are some EXTRACT statistics (ie. data extracted from DB w/ user programs). These tests were also performed w/ OraV7 cache set to 600 blocks. HP cache was set to 200 blocks and no difference was noticed when it was increased to 600 blocks.

       HP 847    DEC 3000/500      Conclusions            Notes
       =======   ============      ===============        =============
real    10.4 s       5.5 s        DEC   89% faster        total CPU time + I/O
user     7.3 s       4.4 s        DEC   66% faster        CPU time in non sys calls
sys      0.5 s       0.4 s        DEC   25% faster        CPU time in sys calls

Conclusions:

• The Alpha 3000/500 does run about twice as fast as the HP 847. Floating point spec marks are about 120 for the Alpha and 60 for the HP. Our data suggests that the spec mark ratings are accurate.
• Loading data into a V7 database degrades the expected performance by > 50%. We believe that this is due to increased constraint checking imposed by the Oracle V7 database. This is supported by the fact that our V6 (user written) loader application compiled to ~400Kb under V6 but compiled to ~3Mb under V7.
• Even though the DEC 3000/500 performed as advertised we decided NOT to purchase one. The reasons were:
  1. Ratio of price/performance no better than an upgrade to our existing HP.
  2. Unavailability of other strategic software and lack of comittment to port by those vendors (especially SAS). This, coupled with the immaturity of OSF/1, posed too great a risk for a production operation.