Description

This project demonstrates performance problems, when using the SwiftKueryMySQL library with multiple simultaneous requests. These problems only show when running on Linux. Mac seems mostly unaffected.

Setup

In order to work, the application needs to be able to connect to any MySQL database, but no tables are needed. The requests simulates a 300 ms query using SELECT SLEEP(0.3). Query time can be tuned by changing the queryDelay variable in TestRoutes.swift:

let queryDelay = 0.3

The username, password, hostname, schema and port number can be changed in the TestRoutes.swift file.

let host = "REPLACE_WITH_HOST_NAME"
let user = "REPLACE_WITH_USERNAME"
let pass = "REPLACE_WITH_PASSWORD"
let schema = "REPLACE_WITH_DATABASE_NAME"
let port = 3306

Compiling and Running on Linux

The following command will build and launch the k-bug application.

bash build-linux.sh

Note that the executable is saved to .build-linux/debug/k-bug.

Compiling and Running on Mac

The problem only seems to manifest when running on Linux. However, should you still want to run on Mac, you may generate an xcode project with all the right settings using the following command:

bash create-xcode-project.sh

You may then open the xcode project, build and run.

Using k-bug

Once you have the k-bug application running on a Linux box, use Chrome to open it http://localhost:8080

I recommend using Chrome, I did not test it with any other browser.

The main page will let you indicate how many simultaneous requests you want to make and what kind.

Use Semaphores: Access the database using the library with semaphores to wait for each completion closure to finish.
Do not use Semaphores: Access the database using the library with nested completion closures.
Mixed: Half of the request will use semaphores and the other half will not.

Column Description

Column Name	Description
Request #	Number of request.
Semaphores	Indicates if the request used semaphores or not.
Processing Duration	How much time was actually used to process the request, does not count query time (300 ms).
Query Duration	Time to execute the query.
Total Duration	Processing Duration + Query Duration.
Start Time	Indicates when the request started processing.
End Time	Indicates when the request finished processing.

What to Look For

The problem is much more evident when using the semaphore requests. Although it also occurs (in a much less obvious way), using the non-semaphore requests.

For example, the semaphore requests may take 4 seconds to finish. That is a lot!

While using the non-semaphore requests some take 20ms while others took 300ms. It may not seem much (specially compared to the semaphore version), but those that took 300 ms means they took 15 times the normal time! This does not happen when running on Mac (for the most part).

It is worth noting that sometimes the request does not even begin processing on time. It seems other requests may be blocking them. This can be seen by looking at the start times. Ideally they should be all almost the same, but sometimes they differ by seconds. By using Chrome network debug pane this requests may show as if they took 4 seconds, but when looking at the processing time it may be 1 second. This means the request was on hold and started 3 seconds later. Is SwiftKuery blocking the Kitura thread!?

Workaround

There is a change that makes all the problems dissappear. Chaning all the calls to DispatchQueue.global().async by Dispatch.global().sync makes it all go away. There are six of this calls: 5 in MySQLConnection.swift and the other in MySQLResultFetcher.swift

I know, I know, this is not how the library is suppose to work... nevertheless, it does work.

kilnerm / k-bug