Compaction cleanupIfNecessary costly when many files in data dir

From what I can tell SSTableWriter.cleanupIfNecessary seems increasingly costly as the number of files in the data dir increases.
It calls SSTable.componentsFor(descriptor, Descriptor.TempState.TEMP) which lists all files in the data dir to find matching components.

Am I roughly correct that (cleanupCost = SSTable count * data dir size)?

We had been doing write load testing with default compaction throttling (16MB/s) and LeveledCompaction.
Unfortunately we haven't been keeping tabs on sstable counts and it grew out of control.

On a system with 300,000 sstables here is an example of our compaction rate. Note that as you're probably aware cleanupIfNecessary is included in the timing:

INFO [CompactionExecutor:48] 2011-11-25 22:25:30,353 CompactionTask.java (line 213) Compacted to [/data1/cassandra/data/MA_DDR/indexes_03-hc-5369-Data.db,]. 5,821,590 to 5,306,354 (~91% of original) bytes for 123 keys at 0.163755MB/s. Time: 30,903ms.

Here's a slightly larger one:
INFO [CompactionExecutor:43] 2011-11-25 22:23:28,956 CompactionTask.java (line 213) Compacted to [/data1/cassandra/data/MA_DDR/indexes_03-hc-5336-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5337-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5338-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5339-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5340-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5341-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5342-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5343-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5344-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5345-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5346-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5347-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5348-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5349-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5350-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5351-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5352-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5353-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5354-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5355-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5356-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5357-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5358-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5359-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5360-Data.db,/data1/cassandra/data/MA_DDR/indexes_03-hc-5361-Data.db,]. 140,706,512 to 137,990,868 (~98% of original) bytes for 2,181 keys at 0.338627MB/s. Time: 388,623ms.

This is with compaction throttling set to 0 (Off).

So I believe because of this it's going to take a very long time to recover from having so many small sstables.
It might be notable that we're using Solaris 10, possibly listFiles() is faster on other platforms?

Is it feasible to keep track of the temp files and just delete them rather than searching for them for each SSTable using SSTable.componentsFor()?

Here's the stack trace for the CompactionExecutor:14 thread that appears to be occupying the majority of the cpu time on this node:

Name: CompactionExecutor:14
State: RUNNABLE
Total blocked: 3 Total waited: 1,610,714

Stack trace:
java.io.UnixFileSystem.getBooleanAttributes0(Native Method)
java.io.UnixFileSystem.getBooleanAttributes(Unknown Source)
java.io.File.isDirectory(Unknown Source)
org.apache.cassandra.io.sstable.SSTable$3.accept(SSTable.java:204)
java.io.File.listFiles(Unknown Source)
org.apache.cassandra.io.sstable.SSTable.componentsFor(SSTable.java:200)
org.apache.cassandra.io.sstable.SSTableWriter.cleanupIfNecessary(SSTableWriter.java:289)
org.apache.cassandra.db.compaction.CompactionTask.execute(CompactionTask.java:189)
org.apache.cassandra.db.compaction.LeveledCompactionTask.execute(LeveledCompactionTask.java:57)
org.apache.cassandra.db.compaction.CompactionManager$1.call(CompactionManager.java:134)
org.apache.cassandra.db.compaction.CompactionManager$1.call(CompactionManager.java:114)
java.util.concurrent.FutureTask$Sync.innerRun(Unknown Source)
java.util.concurrent.FutureTask.run(Unknown Source)
java.util.concurrent.ThreadPoolExecutor$Worker.runTask(Unknown Source)
java.util.concurrent.ThreadPoolExecutor$Worker.run(Unknown Source)
java.lang.Thread.run(Unknown Source)

No matter where I click in the busy Compaction thread timeline in YourKit it's in Running state and showing this above trace, except for short periods of time where it's actually compacting

Thanks,
Eric