mm_cramp_kernel

Leader

Ding Rui @ IBM CDL

author

target

Memory Management ( in kernel )

invalidity

MM(memmory management) in kernel may have bug with heavy load IO operation.

symptom

MM(memmory management) may not handle IO operation failure properly.

When the system had to handle heavy IO operation ,the memory and swap space which were used for caching or buffering was eat up gradually, MM in kernel can not reclaim the occupied memory immediately, system crashed at last because out of memory.

detection strategy

Use our tool to overdo DIO operation in user space.

After a while,long or maybe short,it depends on your real memory sapce on the machine, if the system was crashed or after terminating the tool ,mem and swap never go back to normal,there is the problem.

comment from author

Before anything else, all my words and experiment in this project are my personal opinions.

The IO operation,especially DIO(Direct IO) operation, could "cramp the style" of the kernel when it was overdone. It seems that this problem was hang in doubt for many versions of linux kernel.So it should be doubted.

Our tool reads,writes and truncates file continuously with DIO support.After a while running, the kernel was cramped,system was beginning to kill process and crashed at last because of out of memory. When we terminate the tool before the system go down,the mem and swap nerver go back to normal.

But if we limit the time of IO operations in our tool,after we terminated it,system mem and swap can go back to normal.The system would not crash. It proved that when we overdo DIO operation,kernel may be cramped.

tools

mm_cramp_kernel_1.0 is now available.

measurement results

history