This experiment presents results of an Darshan I/O analysis performed on the CalculiX IO kernel buffered version.
Overall the kernel transferred 40.1 MiB data at 1356.26 MiB/s. The runtime is 0.726 seconds. Compared to the original master version this is a speedup of 4.756.
The following figure shows the average I/O cost per process.
|Darshan reporting the average I/O cost per process of the CalculiX IO buffered version.|
The kernel spends only some percent of its runtime inside write operations. Nearly the whole runtime is spent in the application specific code. In the original master version of this kernel 60% of the runtime was spent with writing data.
Darshan is also able to show the access sizes of I/O operations in the following figure.
|Darshan reporting the most common access sizes of the CalculiX IO buffered version.|
In this case more than 5000 write operations have an access size of 7930 byte. This means that each individual write operation writes almost 8 KB of data. Compard to the original master version this is a huge improvement. Just by using buffered I/O operations we reduced the number of writes from over 1.6 million to just over 5000. This is a factor of 350 because individual writes now also write 350 times more data then in the master version.