You can control different aspects of how an event session behaves by setting the event session options as part of the CREATE EVENT SESSION DDL. The default settings for the event session options are designed to handle most of the common event collection situations so I generally recommend that you just use the defaults. Like everything in the real world though, there are going to be a handful of “special cases” that require something different. This post focuses on identifying the special cases and the correct use of the options to accommodate those cases.
There is a reason it’s called Default
The default session options specify a total event buffer size of 4 MB with a 30 second latency. Translating this into human terms; this means that our default behavior is that the system will start processing events from the event buffer when we reach about 1.3 MB of events or after 30 seconds, which ever comes first.
|Aside: What’s up with the 1.3 MB, I thought you said the buffer was 4 MB?|
The Extended Events engine takes the total buffer size specified by MAX_MEMORY (4MB by default) and divides it into 3 equally sized buffers. This is done so that a session can be publishing events to one buffer while other buffers are being processed. There are always at least three buffers; how to get more than three is covered later.
Using this configuration, the Extended Events engine can “keep up” with most event sessions on standard workloads. Why is this? The fact is that most events are small, really small; on the order of a couple hundred bytes. Even when you start considering events that carry dynamically sized data (eg. binary, text, etc.) or adding actions that collect additional data, the total size of the event is still likely to be pretty small. This means that each buffer can likely hold thousands of events before it has to be processed. When the event buffers are finally processed there is an economy of scale achieved since most targets support bulk processing of the events so they are processed at the buffer level rather than the individual event level. When all this is working together it’s more likely that a full buffer will be processed and put back into the ready queue before the remaining buffers (remember, there are at least three) are full.
I know what you’re going to say: “My server is exceptional! My workload is so massive it defies categorization!” OK, maybe you weren’t going to say that exactly, but you were probably thinking it. The point is that there are situations that won’t be covered by the Default, but that’s a good place to start and this post assumes you’ve started there so that you have something to look at in order to determine if you do have a special case that needs different settings. So let’s get to the special cases…
What event just fired?! How about now?! Now?!
If you believe the commercial adage from Heinz Ketchup (Heinz Slow Good Ketchup ad on You Tube), some things are worth the wait. This is not a belief held by most DBAs, particularly DBAs who are looking for an answer to a troubleshooting question fast. If you’re one of these anxious DBAs, or maybe just a Program Manager doing a demo, then 30 seconds might be longer than you’re comfortable waiting. If you find yourself in this situation then consider changing the MAX_DISPATCH_LATENCY option for your event session. This option will force the event buffers to be processed based on your time schedule. This option only makes sense for the asynchronous targets since those are the ones where we allow events to build up in the event buffer – if you’re using one of the synchronous targets this option isn’t relevant.
Avoid forgotten events by increasing your memory
Have you ever had one of those days where you keep forgetting things? That can happen in Extended Events too; we call it dropped events. In order to optimizes for server performance and help ensure that the Extended Events doesn’t block the server if to drop events that can’t be published to a buffer because the buffer is full. You can determine if events are being dropped from a session by querying the dm_xe_sessions DMV and looking at the dropped_event_count field.
|Aside: Should you care if you’re dropping events?|
Maybe not – think about why you’re collecting data in the first place and whether you’re really going to miss a few dropped events. For example, if you’re collecting query duration stats over thousands of executions of a query it won’t make a huge difference to miss a couple executions. Use your best judgment.
If you find that your session is dropping events it means that the event buffer is not large enough to handle the volume of events that are being published. There are two ways to address this problem. First, you could collect fewer events – examine you session to see if you are over collecting. Do you need all the actions you’ve specified? Could you apply a predicate to be more specific about when you fire the event? Assuming the session is defined correctly, the next option is to change the MAX_MEMORY option to a larger number. Picking the right event buffer size might take some trial and error, but a good place to start is with the number of dropped events compared to the number you’ve collected.
|Aside: There are three different behaviors for dropping events that you specify using the EVENT_RETENTION_MODE option. The default is to allow single event loss and you should stick with this setting since it is the best choice for keeping the impact on server performance low.|
You’ll be tempted to use the setting to not lose any events (NO_EVENT_LOSS) – resist this urge since it can result in blocking on the server. If you’re worried that you’re losing events you should be increasing your event buffer memory as described in this section.
Some events are too big to fail
A less common reason for dropping an event is when an event is so large that it can’t fit into the event buffer. Even though most events are going to be small, you might find a condition that occasionally generates a very large event. You can determine if your session is dropping large events by looking at the dm_xe_sessions DMV once again, this time check the largest_event_dropped_size. If this value is larger than the size of your event buffer [remember, the size of your event buffer, by default, is max_memory / 3] then you need a large event buffer.
To specify a large event buffer you set the MAX_EVENT_SIZE option to a value large enough to fit the largest event dropped based on data from the DMV. When you set this option the Extended Events engine will create two buffers of this size to accommodate these large events. As an added bonus (no extra charge) the large event buffer will also be used to store normal events in the cases where the normal event buffers are all full and waiting to be processed. (Note: This is just a side-effect, not the intended use. If you’re dropping many normal events then you should increase your normal event buffer size.)
Partitioning: moving your events to a sub-division
Earlier I alluded to the fact that you can configure your event session to use more than the standard three event buffers – this is called partitioning and is controlled by the MEMORY_PARTITION_MODE option. The result of setting this option is fairly easy to explain, but knowing when to use it is a bit more art than science. First the science…
You can configure partitioning in three ways: None, Per NUMA Node & Per CPU. This specifies the location where sets of event buffers are created with fairly obvious implication. There are rules we follow for sub-dividing the total memory (specified by MAX_MEMORY) between all the event buffers that are specific to the mode used:
None: 3 buffers (fixed)
Node: 3 * number_of_nodes
CPU: 2.5 * number_of_cpus
Here are some examples of what this means for different Node/CPU counts:
2 CPUs, 1 Node
6 CPUs, 2 Node
40 CPUs, 5 Nodes
|Aside: Buffer size on multi-processor computers|
As the number of Nodes or CPUs increases, the size of the event buffer gets smaller because the total memory is sub-divided into more pieces. The defaults will hold up to this for a while since each buffer set is holding events only from the Node or CPU that it is associated with, but at some point the buffers will get too small and you’ll either see events being dropped or you’ll get an error when you create your session because you’re below the minimum buffer size. Increase the MAX_MEMORY setting to an appropriate number for the configuration.
The most likely reason to start partitioning is going to be related to performance. If you notice that running an event session is impacting the performance of your server beyond a reasonably expected level [Yes, there is a reasonably expected level of work required to collect events.] then partitioning might be an answer. Before you partition you might want to check a few other things:
- Is your event retention set to NO_EVENT_LOSS and causing blocking? (I told you not to do this.) Consider changing your event loss mode or increasing memory.
- Are you over collecting and causing more work than necessary? Consider adding predicates to events or removing unnecessary events and actions from your session.
- Are you writing the file target to the same slow disk that you use for TempDB and your other high activity databases? <kidding> <not really> It’s always worth considering the end to end picture – if you’re writing events to a file you can be impacted by I/O, network; all the usual stuff.
Assuming you’ve ruled out the obvious (and not so obvious) issues, there are performance conditions that will be addressed by partitioning. For example, it’s possible to have a successful event session (eg. no dropped events) but still see a performance impact because you have many CPUs all attempting to write to the same free buffer and having to wait in line to finish their work. This is a case where partitioning would relieve the contention between the different CPUs and likely reduce the performance impact cause by the event session.
There is no DMV you can check to find these conditions – sorry – that’s where the art comes in. This is largely a matter of experimentation. On the bright side you probably won’t need to to worry about this level of detail all that often. The performance impact of Extended Events is significantly lower than what you may be used to with SQL Trace. You will likely only care about the impact if you are trying to set up a long running event session that will be part of your everyday workload – sessions used for short term troubleshooting will likely fall into the “reasonably expected impact” category.
Hey buddy – I think you forgot something
OK, there are two options I didn’t cover: STARTUP_STATE & TRACK_CAUSALITY. If you want your event sessions to start automatically when the server starts, set the STARTUP_STATE option to ON. (Now there is only one option I didn’t cover.) I’m going to leave causality for another post since it’s not really related to session behavior, it’s more about event analysis.