VectorWise TPC-H Results Posted
Posted 2011-02-15 at 06:42 AM by rhann
Updated 2011-02-16 at 04:34 AM by rhann (Clarification of who I mean by "we")
Updated 2011-02-16 at 04:34 AM by rhann (Clarification of who I mean by "we")
The 100GB VectorWise TPC-H benchmark has been out for a few days now, and it's already had a bit of attention in spite of the fact that the official media blitz hasn't started yet.
The TPC-H benchmark is the official, audited version of the DBT-3 benchmark I've been writing about, with the addition that it does some updates too.
There are a few interesting things I notice in the results. First and foremost the official results are very much in line with what we were seeing using DBT-3. I know that because one of the things we established with DBT-3 was the elapsed running time varies pretty much linearly with the size of the database. Double the size of the database and you double the running time. Just a glance at the numbers shows the official test system ran about as fast as our test system would with a 100Gb database.
Two things are worth noting about that: firstly we weren't using trick hardware. We put a bit of thought into configuring it and we tried a few things to get good performance, but the machine cost only about £5,000 all-in. And we got performance that would top the TPC-H league table!
Another thing that is worth mentioning is that the benchmark uses absolutely rock-bottom-of-the-line standard ANSI/ISO SQL. There is no special dialect being used to access funny benchmark-specific tweaks or even proprietary capabilities. Third-party technology-neutral tools talking to VectorWise through JDBC or whatever are going to see the kind of performance demonstrated in the benchmark.
Finally, the logical database design used in this benchmark is a more or less obvious design that would be used in conventional transaction processing. It is not a star or a snowflake or any other "dimensional" transformation. It is therefore feasible to take your production database, maybe pruning off some of the irrelevant tables, maybe obfuscating any sensitive columns, and just use it as-is.
Thanks to low-cost hardware, thoroughly standard SQL, and accessible database designs, VectorWise has virtually no barriers to entry.
The TPC-H benchmark is the official, audited version of the DBT-3 benchmark I've been writing about, with the addition that it does some updates too.
There are a few interesting things I notice in the results. First and foremost the official results are very much in line with what we were seeing using DBT-3. I know that because one of the things we established with DBT-3 was the elapsed running time varies pretty much linearly with the size of the database. Double the size of the database and you double the running time. Just a glance at the numbers shows the official test system ran about as fast as our test system would with a 100Gb database.
Two things are worth noting about that: firstly we weren't using trick hardware. We put a bit of thought into configuring it and we tried a few things to get good performance, but the machine cost only about £5,000 all-in. And we got performance that would top the TPC-H league table!
Another thing that is worth mentioning is that the benchmark uses absolutely rock-bottom-of-the-line standard ANSI/ISO SQL. There is no special dialect being used to access funny benchmark-specific tweaks or even proprietary capabilities. Third-party technology-neutral tools talking to VectorWise through JDBC or whatever are going to see the kind of performance demonstrated in the benchmark.
Finally, the logical database design used in this benchmark is a more or less obvious design that would be used in conventional transaction processing. It is not a star or a snowflake or any other "dimensional" transformation. It is therefore feasible to take your production database, maybe pruning off some of the irrelevant tables, maybe obfuscating any sensitive columns, and just use it as-is.
Thanks to low-cost hardware, thoroughly standard SQL, and accessible database designs, VectorWise has virtually no barriers to entry.








