Description
RAM Prediction
Please refer to the following link for details:
http://www.ccs.neu.edu/home/kapil/courses/cs5600f17/hw4.html
RAM predictor has been implemented by using the following approach.
Data structure:
Each memory block is represented by struct addressList. In Each block we save
the address of the next allocated block and the time taken to write data in the
block through memset. Head points to the first block in the addressList.
Steps in each iteration
1. we allocate a memory block of 4MB which is added to the linked list of
allocated Addresses.
2. we re-write all the allocated addresses from the list of allocated addresses
including the newly allocated address and count the number of blocks that
took double time to write as compared to the previous write time.
The assumption is that such blocks are good candidates for page faults. This
count essentially represents number of page faults in each iteration.
3. We keep a global hash Map, which essentially captures the frequency
distribution of the number of page faults. The index in array represents the
number of page faults occurred and value represents the frequency of it
so far across all iterations. In each iteration we update this hash Map.
4. We keep a sliding window of size 20 to keep track of the number of page
faults happened in last 20 iterations. Calculate the average number of page
faults happening in the last 20 iterations
5. check if the average number of page faults > 20 times the most frequently
occurring page fault. This means that now the number of page faults is 20
times larger than the commonly observed page fault value. And it gives us good
estimate of possibility of thrashing. Program breaks after this point and the
memory allocated so far is the predicted RAM Size.
Observations:
The program was run on a system with 4GB ram (ubuntu) and it consistently
predicted RAM size 3.8 GB Approx.
