OS Spring’04 Virtual Memory: Page Replacement Operating Systems Spring 2004

OS Spring’04 Realizing Virtual Memory  Hardware support Memory Management Unit (MMU): address translation, bits, interrupts  Operating system support Page replacement policy Resident set management Load control  degree of multiprogramming

OS Spring’04 Page Replacement Policy  Resident set maintenance Fixed or variable allocation Per-process or global replacement  Page replacement problem A fixed number of frames, M, is used to map the process virtual memory pages Which page should be replaced when a page fault occurs and all M frames are occupied?

OS Spring’04 Requirements and Metrics  Workload: a sequence of virtual memory references (page numbers)  Page fault rate = #page faults/#memory references  Minimize the page fault rate for workloads obeying the principle of locality Keep hardware/software overhead as small as possible

OS Spring’04 Algorithms  Optimal (OPT)  Least Recently Used (LRU)  First-In-First-Out (FIFO)  Clock

OS Spring’04 Optimal Policy (OPT)  Replace the page which will be referenced again in the most remote future  Impossible to implement Why?  Serves as a baseline for other algorithms

OS Spring’04 Least Recently Used (LRU)  Replace the page that has not been referenced for the longest time  The best approximation of OPT for the locality constrained workloads  Possible to implement  Infeasible as the overhead is high Why?

OS Spring’04 First-In-First-Out (FIFO)  Page frames are organized in a circular buffer with a roving pointer  Pages are replaced in round-robin style When page fault occur, replace the page to which the pointer points to  Simple to implement, low overhead  High page fault rate, prone to anomalous behavior

OS Spring’04 Clock (second chance)  Similar to FIFO but takes page usage into account Circular buffer + page use bit When a page is referenced: set use_bit=1 When a page fault occur: For each page: if use_bit==1: give page a second chance: use_bit=0; continue scan; if use_bit==0: replace the page

OS Spring’04 Example: Page 727 is needed n... Page 9 use = 1 Page 19 use = 1 Page 1 use = 0 Page 45 use = 1 Page 191 use = 1 Page 556 use = 0 Page 13 use = 0 Page 67 use = 1 Page 33 use = 1 Page 222 use = 0 next frame pointer

OS Spring’04 After replacement n... Page 9 use = 1 Page 19 use = 1 Page 1 use = 0 Page 45 use = 0 Page 191 use = 0 Page 727 use = 0 Page 13 use = 0 Page 67 use = 1 Page 33 use = 1 Page 222 use = 0 next frame pointer

OS Spring’04 Example of all algorithms

OS Spring’04 LRU and non-local workloads  Workload: … Typical for array based applications  What is the page fault rate for M=1, …,5?  A possible alternative is to use a Most Recently Use (MRU) replacement policy

OS Spring’04 Belady ’ s Anomaly  It is reasonable to expect that regardless of a workload, the number of page faults should not increase if we add more frames: not true for the FIFO policy: 1, 2, 3, 4, 1, 2, 5, 1, 2, 3, 4,

OS Spring’04 Algorithm comparison

OS Spring’04 Clock algorithm with 2 bits  Use “ modified ” bit to evict unmodified (clean) pages in preference over modified (dirty) pages  Four classes: u=0; m=0: not recently used, clean u=0; m=1: not recently used, dirty u=1; m=0: recently used, clean u=1; m=1: recently used, dirty

OS Spring’04  First scan: look for (0,0) frame, do not change the use bit If (0,0) frame is found, replace it  Second scan: look for (0,1) frame, set use bit to 0 in each frame bypassed If (0,1) frame is found, replace it  If all failed, repeat the above procedure this time we will certainly find something Clock algorithm with 2 bits

OS Spring’04 Page buffering  Evicted pages are kept on two lists: free and modified page lists  Pages are read into the frames on the free page list  Pages are written to disk in large chunks from the modified page list  If an evicted page is referenced, and it is still on one of the lists, it is made valid at a very low cost

OS Spring’04 Page Buffering B B B B B 36N 21N 3N 78N 2N 47N 22N 39N 4N 8N 55N B B B B 36N 21N 3N 78N 2N 47N 22B 39N 4N 8N Page fault: 55 is needed 22 is evicted Buffered frames (B) Normal frames (N)

OS Spring’04 Resident set management  With multiprogramming, a fixed number of memory frames are shared among multiple processes  How should the frames be partitioned among the active processes?  Resident set is the set of process pages currently allocated to the memory frames

OS Spring’04 Global page replacement  All memory frames are candidates for page eviction A faulting process may evict a page of other process  Automatically adjusts process sizes to their current needs  Problem: can steal frames from “ wrong ” processes Leads to thrashing

OS Spring’04 Local page replacement  Only the memory frames of a faulting process are candidates for replacement  Dynamically adjust the process allocation Working set model Page-Fault Frequency (PFF) algorithm

OS Spring’04 The working set model [Denning ’ 68]  Working set is the set of pages in the most recent page references  Working set is an approximation of the program locality

OS Spring’04 The working set strategy  Monitor the working set for each currently active process  Adjust the number of pages assigned to each process according to its working set size  Monitoring working set is impractical  The optimal value of is unknown and would vary

OS Spring’04 Page-Fault Frequency (PFF)  Approximate the page-fault frequency: Count all memory references for each active process When a page fault occurs, compare the current counter value with the previous page fault counter value for the faulting process If < F, expand the WS; Otherwise, shrink the WS by discarding pages with use_bit==0

OS Spring’04 Swapping  If a faulting process cannot expand its working set (all frames are occupied), some process should be swapped out  The decision to swap processes in/out is the responsibility of the long/medium term scheduler  Another reason: not enough memory to run a new process

OS Spring’04 Long (medium) term scheduling  Controls multiprogramming level  Decision of which processes to swap out/in is based on CPU usage (I/O bound vs. CPU bound) Page fault rate Priority Size Blocked vs. running

OS Spring’04 UNIX process states running user running kernel ready user ready kernel blocked zombie sys. call interrupt schedule created return terminated wait for event event done schedule preempt interrupt ready swapped blocked swapped Swap out event done Swap out Swap in

OS Spring’04 Next: File system, disks, etc