File system call performance is just half of vanilla QEMU/KVM one.

solemnify commented 3 years ago

All file system related system call performance is just half of the one in vanilla QEMU/KVM.

All I/O request are forwarded to master node of the GiantVM (refer to the GiantVM paper)
kernel data structure related to file system, such as page cache, dentry cache, inode.....
synchronization overhead, such as spin lock, semaphore...
many core file system scalability issues are largely overapped with the GiantVM file system scalability issues.

Any kernel data structure or synchroniztion primitives are shared on top of the slow DSM layer. Improving DSM performance will give positive impact to this problem...

Our microbecnch for file system performance is the FxMark.

DongDongJu commented 3 years ago

Hello, I am interested in the file system performance in GiantVM. For checking that, I think we need two kinds of benchmarking micro-benchmark and macro-benchmark.

Micro-benchmarking is already satisfied with our FxMark result. How about using RocksDB based benchmark for macro benchmark?

solemnify commented 3 years ago

Hello,

It's very good to see your interest in our SW. It would be very interesting. We have not tried the RocksDB yet though. If you need the result of the FxMark, we will share it soon.

ememos / GiantVM

File system call performance is just half of vanilla QEMU/KVM one. #3