dBatch

This project takes the concept of c-blake's batch. By decreasing the times switching between kernel and user mode, we expected to increase the performance by batching system calls.

Build

build kernel module batch

$ cd module
$ sudo ./build
$ sudo insmod batch.ko

build pre-load library

$ cd user
$ make

Usage

In dBatch, it's easier to use. The only thing need to do are adding batch_start and batch_flush which hint kernel where to start batching.

Example:

#include <linux/batch.h>

int main(int argc, char **argv) {
    batch_start();
    int fd = open("test.txt", O_CREAT | O_WRONLY,
                   S_IWUSR | S_IRUSR | S_IRGRP | S_IWGRP | S_IROTH);
    write(fd, "hihi", 4);
    write(1, "hihi", 4);
    close(fd);
    batch_flush();
    return 0;
}

Application is needed to be compiled with link flag lpthread and be executed with preload lib as prefix:

$ LD_PRELOAD=$PWD/preload.so ./test

More examples can be found under /example.

Feature

Support interleaved non-syscall between syscalls
Support loop
Support multi-thread batching (not portable, only for POSIX & linux glibc)

Performance

So far, we only did simple experiment for measuring performance. Following is summary of experiment, we did 60 times experiment as x-axis. In each iteration, we did 60 times consecutive write in three methods.

Although mode transition is cheaper than context switch, it still has overhead. Optimization of dBatch comes from less time to do mode transition. Mode transition has direct penalty and indirect penalty. The former is consist of executing the trap handler which copies the arguments from the registers to the kernel stack. The later is consist of TLB and cache miss.

	IPC	L1-dcache-load-misses	dTLB-load-misses	cache-misses	page-faults
Normal	0.42	5,8851	33	1,1330	55
dBatch	2.16	3167	17	9471	48

Relative experiments can be found under /experiment

Relative project

c-blake - batch

foxhoundsk / dBatch