Performance on pgbench tpcb with subtransactions is now slightly better than master. See the picture 'savepoints2'. This was achieved by removing unnecessary exclusive locking on CSNLogControlLock in SubTransSetParent. After that change, both versions are mostly waiting on XidGenLock in GetNewTransactionId.
Performance on default pgbench tpcb is also improved. At scale 500, csn is at best 30% faster than master, see the picture 'tpcb500'. These improvements are due to slight optimizations of GetSnapshotData and refreshing RecentGlobalXmin less often. At scale 1500, csn is slightly faster at up to 200 clients, but then degrades steadily: see the picture 'tpcb1500'. Nevertheless, CSN-related code paths do not show up in perf profiles or LWLock wait statistics [1]. I think what we are seeing here is again that when some bottlenecks are removed, the fast degradation of LWLocks under contention leads to net drop in performance. With this in mind, I tried running the same benchmarks with patch from Yura Sokolov [2], which should improve LWLock performance on NUMA machines. Indeed, with this patch csn starts outperforming master on all numbers of clients measured, as you can see in the picture 'tpcb1500'. This LWLock change influences the csn a lot more than master, which also suggests that we are observing a superlinear degradation of LWLocks under increasing contention.
These results look promising for me. Could you try benchmarking using more workloads including read-only and mixed mostly-read workloads?
You can try same benchmarks I used in my talk about CSN in pgconf.eu [1] slides 19-25 (and you're welcome to invent more benchmakrs yourself)
Also, I wonder how current version of CSN patch behaves in worst case when we have to scan the table with a lot of unique xid (and correspondingly have to do a lot of csnlog lookups)? See [1] slide 18. This worst case was significant part of my motivation to try "rewrite xid with csn" approach. Please, find simple extension I used to fill table with random xmins in the attachment.