7.1 replication question

I am using the 7.1 engine under SCO with replication activated. Under certain conditions I am experiencing horrendous performance. Specifically, when I run an OLTP transaction involving hundreds of deletes and updates all other threads in my system grind to a halt. Transactions that would normally take 5 seconds to complete take 1.5 minutes. After sleuthing around with onstat -l and onstat -u I noticed that my hog process and the B-tree cleaner thread were constantly showing up as waiting for one or two of the three available logical log buffers. I initially set my logical log buffers to 32K (16 pages) but then kicked it as high as 512K (256 pages) thinking that this might reduce contention for the logs. To no avail: I am getting about 3 log pages per I/O write no matter what I set LOGUBFF to. It is almost as if every transaction is causing the logical log records to flush from the buffer to disk. The only plausible theory I have is that the replication (which I run in non-buffered synchronous mode) is causing an excessive amount of latch hold time for each log buffer as well as causing the entire database to run at the speed of the replication acknowledgement process (constrained to network speeds). At low transaction volumes I wouldn't saturate the replication acknowledgement process but at higher loads I might. Anybody have any experience with tuning replication to work under heavy OLTP loads?