1   Introduction

In this lab you will be writing a dynamic storage allocator for C programs, i.e., your own version of the malloc, free and realloc routines.  You are encouraged to explore the design space creatively and implement an allocator that is correct, efficient, and fast.

2    Logistics

You are expected to work in a group of two people. Any clarifications and revisions to the assign- ment will be posted on the class website and discussion forum.

3   Hand Out Instructions

You will again be using Git for this project. The upstream Gitlab repository is located at https: //git.cs.vt.edu/cs3214-staff/malloclab. To begin, you will fork the repository using the link on Gitlab.  Then, you should set your repository visibility to private in the settings for your fork. Not doing so is a potential honor code violation. You may also give your partner access to the repository in the settings.  From there, both partners may clone the repository using git clone   and begin work as with projects 1 and 2.

Themalloclab directory contains a number of files. The only file you will be creating and handing in is mm.c. The mdriver.c program is a driver program that allows you to evaluate the perfor- mance of your solution. Use the command make to generate the driver code and run it with the command ./mdriver  -V. (The -V flag displays helpful summary information.)

Looking at the file mm.c you’ll notice a C structure team into which you should insert the re- quested identifying information about the students comprising your programming team.  You may choose the team name freely. Please use your VT (@vt.edu) accounts for the email addresses.

Do this right away so you don’tforget.

– if ptr is not NULL, it must have been returned by an earlier call to mm malloc or mm realloc . The call to mm realloc changes the size of the memory block pointed to by ptr (the old block) to size bytes and returns the address of the new block. Notice that the address of the new block might be the same as the old block, or it might be different, depending on your implementation, the amount of internal fragmentation in the old block, and the size of the realloc request.

The contents of the new block are the same as those of the old ptr block, up to the minimum of the old and new sizes. Note that the new size maybe smaller than the old size. Everything else is uninitialized.  For example, if the old block is 8 bytes and the new block is 12 bytes, then the first 8 bytes of the new block are identical to the first 8 bytes of the old block and the last 4 bytes are uninitialized. Similarly, if the old block is 8 bytes and the new block is 4 bytes, then the contents of the new block are identical to the first 4 bytes of the old block.

These semantics match the semantics of the corresponding libcmalloc, realloc, and free routines. Type man  mallocto the shell for complete documentation.

5   Heap Consistency Checker

Dynamic memory allocators are notoriously tricky beasts to program correctly and efficiently. They are difficult to program correctly because they involve untyped pointer manipulation. You may find it very helpful to write a heap checker that scans the heap and checks it for consistency.

Some examples of what a heap checker might check are:

• Is every block in the free list marked as free?

• Are there any contiguous free blocks that somehow escaped coalescing?

• Is every free block actually in the free list?

• Do the pointers in the free list point to valid free blocks?

• Do any allocated blocks overlap?

• Do the pointers in a heap block point to valid heap addresses?

Your heap checker will consist of the function int  mm check(void) in mm.c . It will check any invariants or consistency conditions you consider prudent. It returns a nonzero value if and only if your heap is consistent.  You are not limited to the listed suggestions nor are you required to check all of them. You are encouraged to printout error messages when mm check fails.

This consistency checker is only for your own debugging during development. When you submit mm.c, make sure to remove any calls to mm check as they will slowdown your throughput.

6    Support Routines

The memlib.c package provides the memory system for your dynamic memory allocator. You can invoke the following functions in memlib.c:

• void  *mem sbrk(int  incr): Expands the heap by incr bytes, where incris a positive non-zero integer and returns a generic pointer to the first byte of the newly allocated heap area. The semantics are identical to the Unix sbrk function, except that mem sbrk accepts only a positive non-zero integer argument, and except that mem sbrk returns NULL on failure rather than -1.

• void  *mem heap lo(void): Returns a generic pointer to the first byte in the heap .

• void  *mem heap hi(void): Returns a generic pointer to the last byte in the heap .

•  size t  mem heapsize(void): Returns the current size of the heap in bytes.

•  size t  mem pagesize(void): Returns the system’s page size in bytes (4K on Linux sys- tems).

7   The Trace-driven Driver Program

The driver program mdriver.c in themalloclab distribution tests your mm.c package for cor- rectness, space utilization, and throughput. The driver program is controlled by a set of trace files, examples of which are included in the tar distribution. Each trace file contains a sequence of allo- cate, reallocate, and free directions that instruct the driver to call your mm malloc , mm   realloc, and mm free routines in some sequence. To grade your submission, we will use the trace files in the default directory

/home/courses/cs3214/malloclab/traces.

The driver mdriver.c accepts the following command line arguments:

•  -s: Run the trace files with a 0.75, 1.0, and 1.25 multiplier on the sizes. This will be used for grading.

•  -t  : Look for the default trace files in directory tracedir instead of the de- fault directory defined in config.h. You may try -t  ˜cs3214/malloclab/oldtraces for shorter versions of these traces for faster debugging.

•  -f  : Use one particular tracefile for testing instead of the default set of tracefiles.

•  -h: Print a summary of the command line arguments.

•  -l: Run and measure libcmalloc in addition to the student’smalloc package.

•  -v: Verbose output. Print a performance breakdown for each tracefile in a compact table.

•  -V: More verbose output. Prints additional diagnostic information as each trace file is pro- cessed.  Useful during debugging for determining which trace file is causing your malloc package to fail.

•  -n:  No heap randomization.  This will use a fixed-address memory region on which to simulate the heap. Use this if you need to trackdown corruption of specific addresses, for instance via gdb’s watchpoints.

8   Programming Rules

• You must not change any of the interfaces in mm.h.

• You must not invoke any memory-management related library calls or system calls.  This rule forbids the use of malloc, calloc, free, realloc, sbrk, brk or any variants of these calls in your code. Using these calls would not make sense because this lab asks you to implement their functionality.

• You may define a limited amount of static global variables.

• For consistency with the libcmalloc package, which returns blocks aligned on 16-byte boundaries, your allocator must always return pointers that are aligned to 16-byte bound- aries. The driver will enforce this requirement for you.

• You must not implement a pure implicit list allocator, but an example of such an allocator is provided in the starter code, and you may use it as a starting point.

9   Thread-Safety and Scalability

The driver program supports benchmarking of your allocator with multiple threads (option -m ).  It is easy to make an allocator safe using a global lock, as the provided mm ts.c sketch shows.

For extra credit, try to come up with better strategies to increase the multi-threaded throughput of your allocator.

10   Evaluation

You will receive zero points if you break any of the rules or your code is buggy and crashes the driver. Otherwise, your grade will be calculated as follows:

•  Correctness (40 points). The points are awarded if your solution passes the correctness tests performed by the driver program, regardless of performance, provided you implemented an allocator that is not the provided implicit list allocator.

Minimum Requirement: Passing the correctness portion of the test for all provided traces is a minimum requirement for this project, as stipulated in the syllabus.

•  Single-threaded Performance (40 points).  Two performance metrics will be used to evaluate your solution:

–  Space utilization: The ratio between the peak aggregate amount of memory used by the driver (i.e., allocated via mm malloc or mm realloc but not yet freed via mm free) and the size of the heap used by your allocator.  The optimal ratio equals to 1 — in that case, the heap grew exactly as much as was needed to accommodate the aggre- gate amount of allocated memory when at itspeak.  You should find good policies to minimize fragmentation in order to make this ratio as close as possible to the optimum.

–  Throughput: The average number of operations completed per second.

The driver program summarizes the performance of your allocator by computing a perfor- mance index, P, which is a weighted sum of the space utilization and throughput

where U is your space utilization, T is your throughput, and Topt  is the throughput of an optimized implementation of malloc on our system on the default traces.2 . The performance index favors space utilization over throughput, with a value of w = 0.6.

Observing that both memory and CPU cycles are expensive system resources, we adopt this formula to encourage balanced optimization of both memory utilization and throughput. Ideally, the performance index will reach P = w + (1 - w) = 1 or 100%. Since each metric will contribute at most wand 1 - wto the performance index, respectively, you should not go to extremes to optimize either the memory utilization or the throughput only. To receive a good score, you must achieve a balance between utilization and throughput.

• Documentation, Style, and Revision Control (20 points).

– Your code should be decomposed into functions and use as few global variables as possible.

– Your code should begin with a header comment that describes the structure of your free and allocated blocks, the organization of the free list, and how your allocator ma- nipulatesthe free list. Each function, global or static, should be preceded by a header comment that describes what the function does.

– You should use a minimum of 5 calls to assert.

– You should make proper use of git in your group. This includes periodically checking in milestones in your implementation, and using descriptive log messages.

11   Handin Instructions

Only one team member needs to submit.

12   Hints

•  Use themdriver -f option. During initial development, using tiny trace files will simplify debugging and testing.  We have included two such trace files (short1,2-bal.rep) that you can use for initial debugging.

•  Use themdriver -vand -V options.  The -v option will give you a detailed summary for each trace file.  The -V will also indicate when each trace file is read, which will help you isolate errors.

•  Compile with gcc  -g and use a debugger. A debugger will help you isolate and identify out of bounds memory references. There are two definitions for CFLAGS in the Makefile: choose the one containing ’-g’ for debugging, and the one containing ’-O3’ to benchmark the per- formance of your solution. After changing the Makefile, do make  clean  all.

•  Study the malloc implementation in the textbook.  The textbook has a detailed example of a simple allocator based on an implicit free list.  Don’t start working on your allocator until you understand everything about the simple implicit list allocator. Note, however, that the C structures used in the naive allocator we provide in our version were added by me and are not discussed in the book.  The book advocates the use of macros such as GET or PUT, see http://csapp.cs.cmu.edu/3e/ics3/code/vm/malloc/mm.cThis approach was required when C compilers (notably, gcc) required that macros and direct pointer arithmetic was used to emit well-performing code.  This is no longer the case.  See also “Define suitable C structures” below.

•  Consider edge conditions.  Consider the case that a block that is freed may not have a left or right neighbor.  A possible strategy is to initialize your heap such that it will appear that there are always allocated “fence” blocks to the left and to the right, which means that the above case never arises. These fence block represent boundary tag headers that are marked as in use.

•  Consider small requests. Depending on which strategy you choose, you will need to roundup small requests. Don’t just think about what happens when allocating a block, consider also what you’ll have to do when freeing this block. Freeing the block may include inserting the block into your free list or lists (or other data structure if you implement one), and thus it must be large enough to holdall link elements plus boundary tags (if used). You will need to consider this both when requesting more memory via mem sbrk() and when splitting a block that maybe too large.

•  Encapsulate your pointer arithmetic in static functions, rather than in C preprocessor macros as suggested in the book.  Pointer arithmetic in memory managers is confusing and error-prone because casting is necessary.  You can reduce the complexity significantly by writing static functions for your pointer operations, which minimize and localize these casts.  Perhaps a future edition of the book will use these methods as well, see Railing 2018[1] for some recent progress on this front.