Benchmarks: Some thoughts / Alguns pensamentos

This article is written in English and Portuguese
Este artigo está escrito em Inglês e Português

English version:

Although I have a perfectly clear disclaimer on the right side of the blog, I’d like to start this article by re-enforcing that disclaimer. Everything I’ll write here are my own thoughts and in no way represent my employer view or position. It’s probable also that the ideas presented here may go against some respectable people’s ideas and some established opinions. I’d like to apologize if someone expects something different. Having said this, the ideas presented here are my real belief, based on my work experience, on my readings and many discussions with some very respectable people.

What is a database benchmark?
Database benchmarks are well defined stress tests and usually when we see references to them chances are that it’s all about the TPC council published benchmarks. The TPC council is a non-profit organization whose members are database software vendors, hardware makers, and other IT related organizations. The purpose of the TPC (Transaction Processing Council) is to “define transaction processing and database benchmarks and to disseminate objective, verifiable TPC performance data to the industry”.
The TPC has defined several specs for benchmarks. The specs have designations like:

  • TPC-C
    OLTP benchmark
  • TPC-D
    Datawarehouse benchmark (deprecated)
  • TPC-H
    Datawarehouse benchmark (current)
  • TPC-DS
    Decision support benchmark (new)
  • TPC-E
    OLTP benchmark created to replace the TPC-C

The benchmark specs are very detailed and include everything that defines the tests. This include the database schema, the queries that are to be run, the several rules that specify what can and can’t be used in terms of features, and also what must be used (like referential and check constraints etc.).
The benchmark reports include the measures taken, which are typically specified in number of transactions per time unit (the transactions are well defined in the specs). They also include the total system cost, which should include hardware and software costs including acquisition and support for a well defined period of time (3 years for example)
And these specs have evolved over time, due to several reasons:

  • New hardware and software features turned earlier specs useless (like materialized views which killed TPC-D)
  • New uses of software made some old specs look a bit outdated, meaning that new specs were created to better match the new reality

Personally I have no reason to assume that TPC was not created with good intentions. But for reasons that will be clarified along this article, I give no credit to the benchmarks. They only measure the will of a supplier to win, and how well it can twist the benchmark rules in order to achieve something worth publishing. Please note that although I have some commercial understanding of the market, I’m basically a technician. If you’re talking to top management of big companies, they tend to don’t understand any technical related argument, and they might, because of that, give some credit to the benchmarks published on But if you’re talking to some technically aware person, I believe it’s easy to dismantle a benchmark in around 10m (I’ll try it for the top TPC-C published as an exercise).
Having said this, running a publishable benchmark represents an enormous technical and economic effort, and I must grant credit to the companies which do it.

The TPC-E mystery…
TCP-E benchmark specification was introduced because TPC-C had a lot of holes that made it unrealistic. You can easily find several opinions stating it’s a better benchmark than it’s predecessor.
But if you check the results you’ll see that only one database vendor, Microsoft, has publish results for this. The mystery is why? Of course, MS supporters say it’s because no one can beat them. No one else believes that. Most people who dedicate some time to study this believe that this is a leap frog game and it really depends on the investment. And if it’s still done (or was) with TPC-C, a very mature specification, it should be even easier to do with a relatively new benchmark (the tricks are easier to discover and implement while the specs are still new). I’ve found several references on the Internet speaking about why no one entered the TPC-E “game”, but none of them is conclusive (from my perspective). I can leave some references here for your own research:

You can decide for yourself. Personally I think the issue is related to what Mr. Jerry Keesee (IBM Informix database development director) explained in a public webcast (Chat with the labs) on 29 January 2009. More on that later.
But currently for OLTP, the most popular benchmark is TPC-C. TPC-E apparently is being pushed by hardware vendors (including IBM).

Benchmarks and Informix
Informix was a regular leader of benchmarks in the nineties. It used to partner with hardware vendors to achieve several top result in several categories.
The subject of benchmarks is very sensitive within the Informix community. Many people strongly believe that IBM should run official benchmarks using Informix. IBM never did it after acquiring Informix Software Inc. We may understand or not that position, but the reasons were clearly explained by Mr. Jerry Keesee, in a very clear answer to the question “How come IBM doesn’t participate in public benchmarks of IDS? Like the benchmarks?”. The question was asked by a well known (and particularly critical of IBM) participant of the Informix forums at the end of a webcast in 29 January 2009. The reasons presented were:

  1. IBM has been doing TPC-C benchmarks with DB2 for a very long time. If we do one with Informix only two things can happen and they’re both bad for IBM:
    1. Informix gets a better number, and the competition and analysts would crush us (IBM and DB2)
    2. Informix gets a worse number, and the competition and analysts would crush us (IBM and Informix)
  2. We could consider TPC-E, but currently there’s only one vendor (MS) who published results on this kind of benchmark. Once we publish one (which would be better in absolute numbers or cost, since no one publishes a benchmark that doesn’t show an improvement), we would have entered a very expensive race, because the vendor who is surpassed will probably reply and the leap frog game would start. Jerry prefers to invest on new features and product improvements which directly bring benefits to the customers.

You can’t ask for something clearer than this. Meanwhile a benchmark on MDM (Meter Data Management) was done and published and Informix got a great result but this is not a standard. So it does not satisfy people who really want TPC results.

Very recently you may have noticed that the words “TPC-C” and “Informix” were floating around the social networks and some Informix related sites. That’s because Eric Vercelletto decided to pick the TPC-C schema and specifications and run a non-oficial database stress test. While some people were jumping around in happiness others were criticizing him. My position is much more neutral, and I think most of the people talking about it never took the time to make a deep analysis of a TPC-C benchmark result. The opinions tend to be divided between something like “yeah!!! Informix finally has a TPC-C benchmark. Now we can show the performance Informix can achieve” and “Oh… The result is so low. He used the free version. It’s useless”. Really, if you want to speak about it, let’s spend some time to look at some facts:

  • The benchmark run by Eric is not a true TPC-C benchmark. It’s not official, it’s not audited and yes, the number is very low if you compare it to other official published results (which by the way can’t be done for legal and technical reasons)
  • Yes, Eric used the Innovator-C edition, which is a cut down version free of costs. It has limits and lacks some features that could help (only 2GB of RAM, no partitioning etc.)
  • Eric fought against technical problems with the clients sending the transactions. In fact he put the clients and the database on the same machine. Something you’ll never see in a true benchmark
  • Eric used 4 hard disks. You can find a published result on site with the same number of cores in the database server, that used 200 hard disks. Yes, you read correctly, two hundred hard disks (but for the top results the numbers are on the thousands)
  • The same  published be…

Auteur : (Fernando Nunes)

, , , , , , , , , , , , ,

No comments yet.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.