Compactions are slow

I have noticed that compactions involving a lot of sstables are very slow (for example major compactions). I have attached a cassandra stress profile that can generate such a dataset under ccm. In my local test I have 2567 sstables at 4Mb each.

I added code to track wall clock time of various parts of the code. One problematic part is ManyToOne constructor. Tracing through the code for every partition creating a ManyToOne for all the sstable iterators for each partition. In my local test get a measy 60Kb/sec read speed, and bottlenecked on single core CPU (since this code is single threaded) with it spending 85% of the wall clock time in ManyToOne constructor.

As another datapoint to show its the merge iterator part of the code using the cfstats from https://github.com/instaclustr/cassandra-sstable-tools/ which reads all the sstables but does no merging gets 26Mb/sec read speed.

Tracking back from ManyToOne call I see this in UnfilteredPartitionIterators::merge

                for (int i = 0; i < toMerge.size(); i++)
                {
                    if (toMerge.get(i) == null)
                    {
                        if (null == empty)
                            empty = EmptyIterators.unfilteredRow(metadata, partitionKey, isReverseOrder);
                        toMerge.set(i, empty);
                    }
                }
 

Not sure what purpose of creating these empty rows are. But on a whim I removed all these empty iterators before passing to ManyToOne and then all the wall clock time shifted to CompactionIterator::hasNext() and read speed increased to 1.5Mb/s.

So there are further bottlenecks in this code path it seems, but the first is this ManyToOne and having to build it for every partition read.