Thanks, Miklos. For measuring the performance, bonnie++ was used over passthrough_ll mount on tmpfs. When taking numbers on vm, I could see non-deterministic behaviour in the results. Therefore core binding was used for passthrough_ll and bonnie++, keeping them on separate cores. Here are the google sheets having performance numbers. https://docs.google.com/spreadsheets/d/1JRgF8DTR9xk5zz3_azmLcyy5kW3bgjjItmS8CYsAoT4/edit#gid=0 https://docs.google.com/spreadsheets/d/1JRgF8DTR9xk5zz3_azmLcyy5kW3bgjjItmS8CYsAoT4/edit#gid=1833203226 Following are the libfuse patches(commit on March 7 and March 8 in first link) which were used to test these changes https://github.com/aakefbs/libfuse/commits/atomic-open-and-no-flush https://github.com/libfuse/libfuse/pull/644 Parameters used in mounting passthrough_ll: numactl --localalloc --physcpubind=16-23 passthrough_ll -f -osource=/tmp/source,allow_other,allow_root, cache=never -o max_idle_threads=1 /tmp/dest (Here cache=never results in direct-io on the file) Parameters used in bonnie++: In sheet 0B: numactl --localalloc --physcpubind=0-7 bonnie++ -x 4 -q -s0 -d /tmp/dest/ -n 10:0:0:10 -r 0 -u 0 2>/dev/null in sheet 1B: numactl --localalloc --physcpubind=0-7 bonnie++ -x 4 -q -s0 -d /tmp/dest/ -n 10:1:1:10 -r 0 -u 0 2>/dev/null Additional settings done on the testing machine: cpupower frequency-set -g performance Running bonnie++ gives us results for Create/s, Read/s and Delete/s. Below table summarises the numbers for these three operations. Please note that for read of 0 bytes, bonnie++ does ops in order of create-open, close and stat but no atomic open. Therefore performance results in the sheet 0B had overhead of extra stat calls. Whereas in sheet 1B, we directed bonnie++ to read 1 byte and this triggered atomic open call but numbers for this run involve overhead for read operation itself instead of just plain open/close. Here is the table summarising the performance numbers Table: 0B Sequential | Random Creat/s Read/s Del/s | Creat/s Read/s Del/s Patched Libfuse -3.55% -4.9% -4.43% | -0.4% -1.6% -1.0% Patched Libfuse + No-Flush +22.3% +6% +5.15% | +27.9% +14.5% +2.8% Patched Libfuse + Patched FuseK +22.9% +6.1% +5.3% | +28.3% +14.5% +2.3% Patched Libfuse + Patched FuseK + No-Flush +33.4% -4.4% -3.73% | +38.8% -2.5% -2.0% Table: 1B Sequential | Random Create/s Read/s Del/s | Create/s Read/s Del/s Patched Libfuse -0.22% -0.35% -0.7% | -0.27% -0.78% -2.35% Patched Libfuse + No-Flush +2.5% +2.6% -9.6% | +2.5% -8.6% -6.26% Patched Libfuse + Patched FuseK +1.63% -1.0% -11.45% | +4.48% -6.84% -4.0% Patched Libfuse + Patched FuseK + No-Flush +32.43% +26.61% +076% | +33.2% +14.7% -0.40% Here No-Flush = No flush trigger from fuse kernel into libfuse In Table 1B, we see 4th row has good improvements for both create and Read whereas Del seems to be almost not changed. In Table 0B, 3rd row we have Read perf reduced, it was found out that this was caused by some changes in libfuse. So this was fixed and in Table 1B, same row, we can see increased numbers. In Table 0B, 3rd row, we have good numbers because bonnie++ used 0 bytes to read and this changed behaviour and impacted perf whereas for the same row, Table 1B we have reduced numbers because it involved flush calls for 1 byte from the fuse kernel into libfuse. These changes are not for fuse kernel/users-space context switches only, but our main goal is to have improvement performance for network file systems - Number network round trips - Reduce load on meta servers with thousands of clients Reduced kernel/userspace context switches is 'just' a side effect. Thanks, Dharmendra
