Hi! Are there any efforts to decrease the internal locking in the file/disk IO subsystem, and perhaps make the sector size configurable?
My quest is to get anything in the ballpark of the speeds I get on /dev/rsd0c (~575MB/sec on aligned random 16KB) on ordinary noncached filesystem access (which I experience as capped around 120MB/sec independent of everything, in my tests on a single SATA or NVME SSD). I will need to make additional benchmarks where I try to max two disks concurrently to see if the peak read speed I get is still 120MB/sec then i.e. that is system-wide, however I think that is the case. (First someone suggested in chatroom that the ~120MB/sec cap on all threads on disk IO is because of the absence of multiqueueing support, but then if I understood David Gwyne right, he's saying it's not but instead somehow about the overhead of the IO subsystem itself and the solution to that is general multithread-ization of the code. I was asking myself if overhead also may be due to that the whole filesystem always leads to per-512-byte accesses to the underlying media too. My SSD benchmarks seem to suggest that aligned 16KB reads perform the best both in both random and sequential modes.) Thanks! Mikael
