A) Preliminary analysis
Currently working on the pre-deployment option. A preliminary study was conducted to identify the memory usage associated with the NPB codes as shown in the table. Here RSS is the Resident Set Size which gives the memory usage of the process + shared pages. Vsize represents the total address space of the process. Also it is observed that, with the current configuration, the memory occupied before any process is running on the guest VM is around 250-300 MB. So there probably is scope for pre-deployment. Returning the free pages to Xen before migration is to be investigated and implemented.
| Class B | Class C | |||
| RSS (MB) | Vsize (MB) | RSS (MB) | Vsize(MB) | |
| BT | 107.90 | 119.41 | 406.01 | 423.33 |
| CG | 110.45 | 125.85 | 282.09 | 306.23 |
| EP | 2.26 | 10.61 | 2.28 | 10.61 |
| LU | 59.14 | 50.30 | 183.06 | 193.60 |
| SP | 95.55 | 111.88 | 351.39 | 379.35 |
The migration framework, which is available as a part of the user-land tool suite has been analyzed and the files involved identified. The sequence of execution has been identified and the structure of image file is analyzed. (The flow is available here). Started to instrument the migration framework for adding a function to compare the image/checkpoint files. Currently modified the python framework to suite our requirements and working on the C code now.
C) Gathered details on reading kernel data structures
The plan
is to perform a black box comparison of image files. In case they don’t match a
lot, we need to identify the reason. Inspecting the kernel data structures in
the image file might give us some idea. So some resources have been gathered which
aim at reading the kernel image/dump and interpreting details like which are
user pages, kernel pages, etc,. At this point, we are not sure yet whether this
might be useful or a correct approach.