Toad World® Forums

Memory, Storage, and optimization


#1

So, I am doing a basic load generation. I am not looking for optimizing, but understanding if I am running optimally (figure that question out…) Anyways, this is really and apples to apples test. I have two servers, configured the exact same. So storage and CPU are the same, the only thing that I can tweak is the memory. So, the question is how do I know if I have too much memory or not enough. I am using a TPC-C load with 500 users (scale factor of 1 with 500 users). I am not sure what performance counters I need. Then once I have this “base line”, I have about 20 servers that will be configured the same. Any ideas or help would be great.


#2

I have a few basic TPC-C suggestions; first of all with a TPC-C load with 500 users you will need a scale factor of 50. I would also run the load with varying user loads such as 1, 100, 200, 500, so you can see some performance trends in your graphs.


#3

So, why do I need a scale factor of 50 for 500 users? I don’t entirley understand what one has to do another. How did you know I needed a scale factor of 50 for 500 users?


#4

Take a look at the following link for some good information which counters you should view to monitor SQL Server memory usage.

http://www.sql-server-performance.com/tips/performance_monitor_memory_counter_p1.aspx

Now for the scale factor question, the way the TPC-C is designed there should be a maximum of 10 users per scale factor. The reason for this is that when this limit is exceeded, then more deadlocks can occur due to the way the transactions are written, which is dictated by the TPC-C specification. BMF does not force you to adhere to this limit since you may want to model the OLTP workload with a smaller dataset. Now with that said the main thing is to get a dataset size that is a closer representation, relative to size, to that of the system that will be on the server. The size of your dataset will affect the optimal memory configuration.

I hope this makes sense.

Kevin


#5

Hi, this is very helpful information. ICan assume the same applies to the TCP-E benchmark as well?


#6

Yes everything is applicable to the TPC-E except for the 10 virtual users per scale factor guideline.