GitLab

Uses a more efficient function to load CSV data.

Removed relying of map's mutability

The function uses the fact that rows in trace files are ordered
to make more efficient loading of the traces.

uses flatmap instead of mapvalues and them toList

Having a static DYNCLASSLOADER prevented some garbage collection,
causing memory leaks.

Using the target_type of the metadata was wrong

Had a complicated way of getting the interface type of a collection
class, now the InterfaceName enum type does the conversion.

It is less general, but it's simpler.

Some benchmarks have a lot of duplicates, so we compress the data by
aggregating them, the number of duplicates for each benchmark is then
stored.

Some of the collections are not valid contestants for benchmarking
because they don't have the same semantics.

Error messages were displayed when running the test

Thought this might help with performance (not sure it does)

When two traces are from the same allocation site, with the same target
type and the exact same list of methods, we only keep one.

We used groupby and then processed each group, so we can clear as we go.

The benchmarks didn't use the numberSamples and numberWarmups given
to the constructor

Relied on streams more instead of lists, trying to get faster running
of generated benchmarks

Removed double-generation of benchmarks

There is probably a better way, since some options are then
invalid

The benchmark part was too unstable...

Each traces is loaded, grouped, and benchmarks are created to
try each collection for each benchmarks, then suggestions
are created

Originally, when loading a trace, only one benchmark was created
but there's now a function to also generate benchmarks so that
we can run them and make comparisons

Changed the expected values to fit the new format for benchmark
identifiers

List.of() doesn't exist for java 8

Theres a test for running a benchmark based on a trace
and one from trace benchmarks loaded from a CSV data

The user is able to change the number of iterations
for measurement and warmup in SyntheticBenchmarkRunner

A trace with a different object ID and location should have a different
ID.

We check that the ids are indeed different if we generate
two benchmarks correctly

Removed double iteration over all benchmarks to build the list of
benchmark indexes.

Also fixed a broken switch statement.

Was necessary to enable serialization

The test writes benchmarks to memory and then reads them

Means that deserializing from memory is also possible