The Unbearable Slowness of Some disc_only_copies Transactions

[adam]

We have running Erlang code that is executing about 30 transactions
per second. Each transaction puts a write lock on a couple of records,
then performs a read and a write. We need to hold several hundred
million records simultaneously.

When using disc_copies tables, we can easily manage to run about 160
transactions per second. We also see that each transaction takes an
average of about 750-1400 microseconds to run. Min time is 350
microseconds, and max time seems to go between 30-100 milliseconds.

However, with disc_only_copies tables, individual transaction time
ranges between 500 microseconds up to about 90 seconds(!). Why do
individual transactions sometimes take such a long time with
disc_only_copies?

Thank you!
Adam