Virtual Memory: Page Replacement Operating Systems Fall 2002 OS Fall’02

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 Fall’02

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 Fall’02

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 Fall’02

Algorithms Optimal (OPT) Least Recently Used (LRU) First-In-First-Out (FIFO) Clock OS Fall’02

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 Fall’02

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 Fall’02

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 Fall’02

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 Fall’02

Example: Page 727 is needed 1 2 3 4 5 6 7 8 n . Page 9 use = 1 Page 19 Page 1 use = 0 Page 45 Page 191 Page 556 Page 13 Page 67 Page 33 Page 222 next frame pointer OS Fall’02

After replacement n Page 19 use = 1 Page 9 use = 1 1 Page 1 use = 0 . n Page 19 use = 1 Page 9 use = 1 1 Page 1 use = 0 . . 2 Page 45 use = 0 . next frame pointer Page 191 use = 0 Page 222 use = 0 3 8 Page 727 use = 0 Page 33 use = 1 4 Page 67 use = 1 Page 13 use = 0 7 6 5 OS Fall’02

Example of all algorithms OS Fall’02

LRU and non-local workloads 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 Fall’02

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, 5 1 1 1 1 5 4 4 5 2 2 2 2 1 5 1 3 3 3 3 3 2 2 4 4 4 3 OS Fall’02

Algorithm comparison OS Fall’02

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 Fall’02

Clock algorithm with 2 bits 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 OS Fall’02

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 Fall’02

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 Fall’02

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 Fall’02

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 Fall’02

Approximating the WS Global page replacement All memory frames are candidates for page eviction a faulting process may evict a page of other process Processes with larger WS are expanding whereas those with smaller WS are shrinking Problem: may unjustly reduce the WS of some processes Combine with page buffering OS Fall’02

Approximating the WS Local page replacement Only the memory frames of a faulting process are candidates for replacement Dynamically adjust the process allocation Page-Fault Frequency (PFF) algorithm OS Fall’02

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 Fall’02

Multiprogramming level Too many processes in memory Thrashing, inability to run new processes The solution is swapping: save all the resident set of a process to the disk (swapping out) load the pages of another process instead (swapping in) Long-term and medium term scheduling decides which processes to swap in/out OS Fall’02

Long (medium) term scheduling Decision of which processes to swap out/in is based on The CPU usage Creating a balanced job mix with respect to I/O vs. CPU bound processes Two new process states: Ready swapped Blocked swapped OS Fall’02

UNIX process states running user zombie ready user running kernel schedule sys. call interrupt return ready user interrupt zombie running kernel terminated preempt schedule wait for event ready kernel event done blocked created Swap in Swap out Swap out ready swapped blocked swapped event done OS Fall’02

Segmentation with paging simplifies protection and sharing, enforce modularity, but prone to external fragmentation Paging transparent, eliminates ext. fragmentation, allows for sophisticated memory management Segmentation and paging can be combined OS Fall’02

Address translation Main Memory Paging OS Fall’02 Page Frame Offset Paging Page Table P# + Frame # Seg Table Ptr S # Segmentation Program Segment Table Seg # Page # OS Fall’02

Page size considerations Small page size better approximates locality large page tables inefficient disk transfer Large page size internal fragmentation Most modern architectures support a number of different page sizes a configurable system parameter OS Fall’02

Next: File system, disks, etc OS Fall’02