2,257
edits
Changes
Jump to navigation
Jump to search
→Optimize the pre-dump algorithm
Line 90:
Line 90:
−* It keeps all the memory in pipes, and their number can be huge due to limited one pipe size+* It keeps all the memory in pipes and this memory is unreclaimable for that period+* It [[parasite code|stops and infects]] tasks to drain memory from+
−
−* No pipes, just copy data into temp buffer and send
−* Memory is always reclaimable
−* No infection is needed, just freeze, reset the tracker and proceed
'''Summary:''' Optimize the pre-dump algorithm to avoid pinning to many memory in RAM
'''Summary:''' Optimize the pre-dump algorithm to avoid pinning to many memory in RAM
Current [[CLI/cmd/pre-dump|pre-dump]] mode suffers from several issues
+Current [[CLI/cmd/pre-dump|pre-dump]] mode is used to write task memory contents into image
+files w/o stopping the task for too long. It does this by stopping the task, infecting it and
+draining all the memory into a set of pipes. Then the task is cured and resumed and the pipes'
+contents is written into images (maybe a [[page server]]). This approach creates a big stress
+on memory subsystem, as keeping all the memory in pipes creates a lot of unreclaimable memory
+(pages in pipes are not swappable), as well as the number of pipes themselves can be buge (as
+one pipe doesn't store more than a fixed certain amount of data).
−We can try to use sys_read_process_vm() syscall to mitigate all of the above. To do this we
−need to allocate a temporary buffer in criu, then walk the target process vm by copying the
−memory piece-by-piece into it, then flush the data into image (or page server), then repeat.
−We can try to use sys_read_process_vm() syscall to mitigate all of the above. Benefits:
−Ideally there should be sys_splice_process_vm() syscall in the kernel, that does the same as
Ideally there should be sys_splice_process_vm() syscall in the kernel, that does the same as