After spending some time in the cloud, I figured it is time to come down to earth and start discussing some of the new Auto DOP features some more. As Database Machines (the v2 machine runs Oracle Database 11.2) are effectively selling like hotcakes, it makes some sense to talk about the new parallel features in more detail.
For basic understanding make sure you have read the initial post. The focus there is on Auto DOP and queuing, which is to some extend the focus here. But now I want to discuss the concurrency a little and explain some of the relevant parameters and their impact, specifically in a situation with concurrency on the system.
The idea behind calculating the Automatic Degree of Parallelism is to find the highest possible DOP (ideal DOP) that still scales. In other words, if we were to increase the DOP even more above a certain DOP we would see a tailing off of the performance curve and the resource cost / performance would become less optimal. Therefore the ideal DOP is the best resource/performance point for that statement.
On a normal production system we should see statements running concurrently. On a Database Machine we typically see high concurrency rates, so we need to find a way to deal with both high DOP's and high concurrency.
Queuing is intended to make sure we
Instead of making statements go at a lower DOP we queue them to make sure they will get all the resources they want to run efficiently without trashing the system. The theory - and hopefully - practice is that by giving a statement the optimal DOP the sum of all statements runs faster with queuing than without queuing.
To determine how many statements we will consider running in parallel a single parameter should be looked at. That parameter is called PARALLEL_MIN_TIME_THRESHOLD. The default value is set to 10 seconds. So far there is nothing new here..., but do realize that anything serial (e.g. that stays under the threshold) goes straight into processing as is not considered in the rest of this post.
Now, if you have a system where you have two groups of queries, serial short running and potentially parallel long running ones, you may want to worry only about the long running ones with this parallel statement threshold. As an example, lets assume the short running stuff runs on average between 1 and 15 seconds in serial (and the business is quite happy with that). The long running stuff is in the realm of 1 - 5 minutes.
It might be a good choice to set the threshold to somewhere north of 30 seconds. That way the short running queries all run serial as they do today (if it ain't broken, don't fix it) and allows the long running ones to be evaluated for (higher degrees of) parallelism. This makes sense because the longer running ones are (at least in theory) more interesting to unleash a parallel processing model on and the benefits of running these in parallel are much more significant (again, that is mostly the case).
Now that you know how to control how many of your statements are considered to run in parallel, lets talk about the specific degree of any given statement that will be evaluated. As the initial post describes this is controlled by PARALLEL_DEGREE_LIMIT. This parameter controls the degree on the entire cluster and by default it is CPU (meaning it equals Default DOP).
For the sake of an example, let's say our Default DOP is 32. Looking at our 5 minute queries from the previous paragraph, the limit to 32 means that none of the statements that are evaluated for Auto DOP ever runs at more than DOP of 32.
A basic assumption about running high DOP statements at high concurrency is that you at some point in time (and this is true on any parallel processing platform!) will run into a resource limitation. And yes, you can then buy more hardware (e.g. expand the Database Machine in Oracle's case), but that is not the point of this post...
The goal is to find a balance between the highest possible DOP for each statement and the number of statements running concurrently, but with an emphasis on running each statement at that highest efficiency DOP.
The PARALLEL_SERVER_TARGET parameter is the all important concurrency slider here. Setting this parameter to a higher number means more statements get to run at their maximum parallel degree before queuing kicks in. PARALLEL_SERVER_TARGET is set per instance (so needs to be set to the same value on all 8 nodes in a full rack Database Machine). Just as a side note, this parameter is set in processes, not in DOP, which equates to 4* Default DOP (2 processes for a DOP, default value is 2 * Default DOP, hence a default of 4 * Default DOP).
Let's say we have PARALLEL_SERVER_TARGET set to 128. With our limit set to 32 (the default) we are able to run 4 statements concurrently at the highest DOP possible on this system before we start queuing. If these 4 statements are running, any next statement will be queued.
To run a system at high concurrency the PARALLEL_SERVER_TARGET should be raised from its default to be much closer (start with 60% or so) to PARALLEL_MAX_SERVERS. By using both PARALLEL_SERVER_TARGET and PARALLEL_DEGREE_LIMIT you can control easily how many statements run concurrently at good DOPs without excessive queuing. Because each workload is a little different, it makes sense to plan ahead and look at these parameters and set these based on your requirements.