Computer Systems Lab Stefan M. Freudenberger Computer Systems, ETH - PowerPoint PPT Presentation

Computer Systems Lab Stefan M. Freudenberger Computer Systems, ETH Zürich

Introduction  Course Schedule  Lecture: Monday 4:15 – 5:00 p.m. (RZ F21)  Recitation: TBD  Assignments: in teams, time to be arranged individually  Contact  Stefan Freudenberger, RZ H8.1, stefan.freudenberger@inf.ethz.ch  TBD  Resources  http://www.lst.inf.ethz.ch/teaching/lectures/current/2100/index.php 2011-02-28 Computer Systems Lab 2

Course schedule (tentative) Date Lecture Project 2011-02-21 Introduction 2011-02-28 Memory Allocation Team registration – due 2011-03-03 2011-03-07 Computer Science Colloquium - Prof. Dr. David Padua: And Malloc Lab – due 2011-03-20 the Answer Is: Automatic Parallel Programming. Starts at 16:15 CAB G61 (ends at ca. 17:20) 2011-03-14 2011-03-21 – due 2011-04-03 2011-03-28 2011-04-04 – due 2011-04-17 2011-04-11 Sechseläuten: holiday 2011-04-18 2011-04-25 Easter Monday: holiday 2011-05-02 – due 2011-05-15 2011-05-09 2011-05-16 – due 2011-05-29 2011-05-23 2011-05-30 2011-02-28 Computer Systems Lab 3

Admin Team registration:  E-mail with team members:  Include names, legi numbers, nethz-logins  Subject: “ CSLab: …”  Due: Thursday, March 3  Get team name assignment back  Due: Monday, March 7  Check that you can access your team’s area:  Do this early!  Assignment will be posted on the web site on Monday, March 7  Assignment will be due midnight March 20 2011-02-28 Computer Systems Lab 4

Project 1: Memory Allocation  Summary  Implement a memory allocator: malloc, free, realloc  Goals  Learn how to manage storage allocation  Good exercise to master pointers  Environment  Unix like system  C  Resources  Computer Systems — A Programmer’s Perspective, Chapter 10.9 (Dynamic Memory Allocation), pages 730–755 (1 st Edition). 2011-02-28 Computer Systems Lab 5

The Heap Area of a process’s (virtual) memory User stack that is dynamically allocated  Example (Unix)  Grows upwards Memory mapped region  The brk pointer marks the top of the for shared libraries heap Top of the heap ( brk ) Heap Uninitialized data (.bss) Initialized data (.data) Program (.text) 2011-02-28 Computer Systems Lab 6

Memory Management The memory manager (user space) maintains a process’s heap : a collection of blocks (contiguous chunks of memory) which are either allocated or free  Allocation is explicit ( e.g. , malloc in C or new in Java)  Deallocation can be:  explict ( e.g. , free in C, delete in C++)  implicit (garbage collection in Java, .NET) Issues  The memory manager needs to:  Distinguish block boundaries  Distinguish between free and allocated blocks 2011-02-28 Computer Systems Lab 7

Memory Allocation: Strategies  first-fit: use the first free block which is big enough  best-fit: take smallest fitting block  worst-fit: take biggest available block  quick-fit: best-fit but multiple free-lists (one per block size) 32 8 16 8 16 32 8 8 32 16 8 8 32 16 16 64 8 16 8 32 2011-02-28 Computer Systems Lab 8

Memory Allocation: Bitmaps The simplest (but often slowest) way to manage memory blocks is to mark in a bitmap if a block is free or allocated  Space  For a 4GB memory space and 4KB blocks a 128MB table is necessary to store the free/allocated information  Speed  Slow scan to find a free block 32 8 16 8 0 0 0 0 1 1 1 0 16 32 8 8 0 0 1 1 1 1 0 1 32 16 8 8 1 1 1 1 0 0 0 1 32 16 16 0 0 0 0 0 0 1 1 64 0 0 0 0 0 0 0 0 8 16 8 32 0 1 1 0 1 1 1 1 2011-02-28 Computer Systems Lab 9

Memory Allocation: Dynamic Data Structures  Free blocks are stored in a dynamic data structure  For example, ordered list, ordered tree, …  Lists are usually ordered by allocation address to allow a faster merging of free blocks 32 8 16 8 free list 16 32 8 8 32 16 8 8 32 16 16 64 8 16 8 32 2011-02-28 Computer Systems Lab 10

Memory Allocation: Dynamic Data Structures  Free blocks are stored in a dynamic data structure  For example, ordered list, ordered tree, …  Lists are usually ordered by allocation address to allow a faster merging of free blocks  When two adjacent blocks are both free, they are coalesced into a single block  Initially: 8 16 8  After free: 8 16 8  After coalescing: 32 2011-02-28 Computer Systems Lab 11

Memory Allocation: Lists  Space  Link information as well as some flags ( i.e. , free or allocated) has to be stored in the block  Speed  Depending on the chosen data structure operations can be implemented efficiently 2011-02-28 Computer Systems Lab 12

Block Format Most allocators store information (allocated/free, size, …) in the blocks themselves  Example: size flags size flags ptr link information ptr payload payload padding (optional) padding (optional)  Link information:  implicit : blocks are traversed using the size field only  explicit : link information is embedded in the block 2011-02-28 Computer Systems Lab 13

Block Format (Footers) Implicit allocators have the problem that to perform coalescing they need information about the previous block.  Knuth [1973] suggested to put a copy size flags of the block header at the end of the link information block (footer)  When freeing a block an implicit allocator payload can then check the header of the next block and the footer of the previous one  Footers are not necessary for allocated padding (optional) size flags blocks if we store the allocated/free information in the header of the next block D. E. Knuth. The Art of Computer Programming. Addison Wesley, 1973. Volume 1, Chapter 2, pp. 228–463. 2011-02-28 Computer Systems Lab 14

Memory Allocation: Space Utilization  External fragmentation  Unused space among the blocks – depends on the allocation strategy  Internal fragmentation  Unused space inside the blocks – depends on the possible block sizes (granularity)  Data structures  Space used to maintain data structures cannot be allocated (bitmaps, pointers, flags, …) 2011-02-28 Computer Systems Lab 15

Memory Allocation: Buddy System  An example of allocation strategy:  Each block has a size of 2 k  The system maintains a list for each size  Block allocation:  Find a block which fits  If necessary split a larger block, and put the remaining part in the correct list  Free:  Mark the block as free  If is possible merge with a free neighbor  The buddy system is used in several operating systems (Windows, Linux, Oberon, …). 2011-02-28 Computer Systems Lab 16

Memory Allocation: Buddy System (example)  Free: Free list:  Initial: 8 8 8 8 16 16 16 32 2011-02-28 Computer Systems Lab 17

Memory Allocation: Buddy System (example)  Free: Free list:  Initial: 8 8 8 8 8 16 16 16  Free: 32 8 8 16 2011-02-28 Computer Systems Lab 18

Memory Allocation: Buddy System (example)  Free: Free list:  Initial: 8 8 8 16 16 16 16  Free: 32 8 8 16  Coalesce: 16 16 2011-02-28 Computer Systems Lab 19

Memory Allocation: Buddy System (example)  Free: Free list:  Initial: 8 8 8 16 16  Free: 32 32 8 8 16  Coalesce: 16 16  Coalesce again: 32 2011-02-28 Computer Systems Lab 20

Memory Allocation: Buddy System (example)  Allocate(8): Free list:  Initial: 8 32 16 32 32 2011-02-28 Computer Systems Lab 21

Memory Allocation: Buddy System (example)  Allocate(8): Free list:  Initial: 8 32 16 16 16  Split free block: 32 16 16 2011-02-28 Computer Systems Lab 22

Memory Allocation: Buddy System (example)  Allocate(8): Free list:  Initial: 8 8 8 32 16 16  Split free block: 32 16 16  Split free block again: 8 8 16 2011-02-28 Computer Systems Lab 23

Memory Allocation: Buddy System (example)  Allocate(8): Free list:  Initial: 8 8 32 16 16  Split free block: 32 16 16  Split free block again: 8 8 16  Allocate: 8 8 16 2011-02-28 Computer Systems Lab 24

Memory Allocation: Slab Allocator  Disadvantages of the buddy system:  Internal fragmentation (max 50%)  Bad distribution of the block addresses (bad caches performance)  Solaris introduced the concept of a slab allocator :  Based on the idea that processes are likely to request a lot of objects of the same size: these objects can be kept and reused  Slabs are collections of objects of the same size  Sizes (and addresses) are not geometrically distributed  Used by AmigaOS (4), Linux ( ≥ 2.2) and Solaris (≥ 2.4) J. Bonwick. The Slab Allocator: An Object-Caching Kernel Memory Allocator. Proc. 1994 USENIX Annual Tech. Conf, pp. 87–98, June 1994, Boston, MA. 2011-02-28 Computer Systems Lab 25

Lab: Task  Write a dynamic memory allocator, i.e. , implement:  malloc  free  realloc  Goals of the lab:  Understand how a real memory allocator works  Evaluate different design strategies  A nice exercise for low level C programming 2011-02-28 Computer Systems Lab 26