15-213/18-213/15-513, Spring 2021
Malloc Lab: Writing a Dynamic Storage Allocator Assigned: Tue, Mar 23
This lab requires submitting two versions of your code: one as an initial checkpoint, and the second as your final version. The due dates of each part are indicated in the following table:
Don’t forget that due dates are at 11:59pm Eastern Time. See section 7 for details on how each section is scored.
1 Introduction
In this lab you will write a dynamic memory allocator which will consist of the malloc, free, realloc, and calloc functions. Your goal is to implement an allocator that is correct, efficient, and fast.
We strongly encourage you to start early. The total time you spend designing and debugging can easily eclipse the time you spend coding.
Bugs can be especially pernicious and difficult to track down in an allocator, and you will probably spend a significant amount of time debugging your code. Buggy code will not get any credit.
This lab has been heavily revised from previous versions. Do not rely on advice or information you may find on the Web or from people who have done this lab before. It will most likely be misleading or outright wrong1. Be sure to read all of the documentation carefully and especially study the baseline implementation we have provided.
1Not to mention the fact that it would be an academic integrity violation! 1
Version
Due Date
Max. Grace Days
Last Hand-in Date
Weight in Final Grade
Checkpoint Final
Mar 30 Apr 8
2 2
Apr 2 Apr 11
4% 7%
Contents
1 Introduction 1
2 Logistics 3
3 Required Functions 4
4 Support Routines 6
5 Programming Rules 7
6 Driver Programs 9
6.1 Tracefiles…………………………………… 9 6.2 Command-linearguments …………………………… 9
7 Scoring 11
7.1 Autogradedscore……………………………….. 11 7.1.1 Utilization……………………………….. 11 7.1.2 Throughput………………………………. 12 7.1.3 Performanceindex …………………………… 12 7.1.4 Autogradeddeductions ………………………… 13
7.2 HeapConsistencyChecker…………………………… 13
7.3 Style……………………………………… 14
7.4 HandinInstructions ……………………………… 15
8 Useful Tips 16
9 Office Hours 17
10 Strategic Advice 18
A Performance Evaluation 20
A.1 ApproximateExpectedResultsfromOptimizations. . . . . . . . . . . . . . . . . . . 20 A.2 MachineDependencies…………………………….. 20 A.3 PerformancePoints………………………………. 20
2
B Viewing Heap Contents with GDB 22
B.1 Viewingtheheapwithoutahelperfunction ………………….. 22 B.2 Viewingtheheapwiththehprobehelperfunction . . . . . . . . . . . . . . . . . . . 23
3
2 Logistics
This is an individual project. You should do this lab on one of the Shark machines.
To get your lab materials, click “Download Handout” on Autolab, enter your Andrew ID, and
follow the instructions. Then, clone your repository on a Shark machine by running:
$ git clone https://github.com/cmu15213s21/malloclab-s21-
The only file you will turn in is mm.c, which contains your solution. The provided code allows you to evaluate your implementations. Using the command make will generate four driver programs: mdriver, mdriver-dbg, mdriver-emulate, and mdriver-uninit, as described in section 6. Your final autograded score is computed by driver.pl, as described in section 7.1.
To test your code for the checkpoint submission, run mdriver and/or driver.pl with the -C flag. To test your code for the final submission, you don’t need any additional flags.
You can get some idea of your program performance and the resulting value you will get for the throughput portion of your grade. However, the Autolab servers will generate different throughput values, and these will determine your actual score. This is discussed in more detail in Section 7.
4
3 Required Functions
Your allocator will implement the following functions, declared in mm.h and defined in mm.c:
bool mm_init(void);
void *malloc(size_t size);
void free(void *ptr);
void *realloc(void *ptr, size_t size);
void *calloc(size_t nmemb, size_t size);
bool mm_checkheap(int);
We provide you two versions of memory allocators:
mm.c: A fully-functional implicit-list allocator. We recommend that you use this code as your starting point.
Note that the provided code does not implement coalesce block(). Although it is a valid memory allocator, it will result in a high amount of external fragmentation, so you should implement coalescing. We strongly recommend considering all cases you need to implement before writing code for this function; the lecture slides should help you identify and reason about these cases.
mm-naive.c: A functional implementation that runs quickly but gets very poor utilization, because it never reuses any blocks of memory.
Your allocator must run correctly on a 64-bit machine. It must support a full 64-bit address space, even though current implementations of x86-64 machines support only a 48-bit address space.
Your submitted mm.c must implement the following functions:
• bool mm_init(void): Performs any necessary initializations, such as allocating the initial heap area. The return value should be false if there was a problem in performing the initialization, true otherwise. You must reinitialize all of your data structures in this function, because the drivers call your mm init function every time they begin a new trace to reset to an empty heap.
• void *malloc(size_t size): The malloc routine returns a pointer to an allocated block payload of at least size bytes. The entire allocated block should lie within the heap region and should not overlap with any other allocated block.
Your malloc implementation must always return 16-byte aligned pointers.
• void free(void *ptr): The free routine frees the block pointed to by ptr. It returns nothing. This routine is only guaranteed to work when the passed pointer was returned by an earlier call to malloc, calloc, or realloc and has not yet been freed. free(NULL) has no effect.
5
• void *realloc(void *ptr, size_t size): The realloc routine returns a pointer to an allocated region of at least size bytes with the following constraints:
– if ptr is NULL, the call is equivalent to malloc(size);
– if size is equal to zero, the call is equivalent to free(ptr) and should return NULL;
– if ptr is not NULL, the call is equivalent to free(ptr) followed by malloc(size), except the contents of the new block will be the same as those of the old block, up to the minimum of the old and new sizes.
Your realloc implementation will have only minimal impact on measured throughput or utilization. A correct, simple implementation will suffice.
• void *calloc(size_t nmemb, size_t size): Allocates memory for an array of nmemb ele- ments of size bytes each and returns a pointer to the allocated memory. The memory is set to zero before returning.
Your calloc will not be graded on throughput or performance. A correct, simple implemen- tation will suffice.
• bool mm_checkheap(int line): The mm checkheap function (the heap consistency checker, or simply heap checker) scans the heap and checks it for possible errors.
A quality heap checker is essential for debugging your malloc implementation. Many malloc bugs are too subtle to debug using conventional gdb techniques. A heap consistency checker can help you isolate the specific operation that causes your heap to become inconsistent.
Because of the importance of the consistency checker, it will be graded; section 7.2 describes the requirements for your implementation in greater detail. We may also require you to write your checkheap function before coming to office hours.
The mm checkheap function takes a single integer argument that you can use any way you want. One technique is to use this this argument to pass in the line number where it was called, using the __LINE__ macro:
mm_checkheap(__LINE__);
This allows you to print the line number where mm checkheap was called, if you detect a problem with the heap. Note, however, that mm checkheap will also be called with an argument of 0 by the driver.
The semantics for malloc, realloc, calloc, and free match those of the corresponding libc routines. Type man malloc in the shell for more documentation.
6
4 Support Routines
The memlib.c package simulates the memory system for your dynamic memory allocator. You can invoke the following functions in memlib.c:
• void *mem sbrk(intptr t incr): Expands the heap by incr bytes, where incr is a non- negative 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 will fail for negative arguments. (Data type intptr t is defined to be a signed integer large enough to hold a pointer. On our machines it is 64-bits long.)
• void *mem heap lo(void): Returns a generic pointer to the first valid byte in the heap.
• void *mem heap hi(void): Returns a generic pointer to the last valid byte in the heap.
Careful: the definition of mem heap hi() given here is correct, but it may not be intuitive! If your heap is 8 bytes large, then the “last valid byte” will be 7 bytes from the start — and its address will not be aligned.
• size t mem heapsize(void): Returns the current size of the heap in bytes.
You are also allowed to use the following libc library functions: memcpy, memset, printf, fprintf, and sprintf. Other than these functions and the support routines, your mm.c code may not call any externally-defined function.
7
5
•
•
•
Programming Rules
Any allocator that attempts to explicitly detect which trace is running will receive a penalty of 20 points. On the other hand, you should feel free to write an adaptive allocator—one that dynamically tunes itself according to the general characteristics of the different traces.
You should not change any of the interfaces in mm.h, and your program must compile with the provided Makefile. However, we strongly encourage you to use static helper functions in mm.c to break up your code into small, easy-to-understand segments.
You are not allowed to define any large global data structures such as large arrays, trees, or lists in your mm.c program. However, you are allowed to declare small global arrays, structs and scalar variables such as integers, floats, and pointers in mm.c. In total, your global data should sum to at most 128 bytes. Global values defined with the const qualifier are not counted in the 128 bytes. This will be tested by automated means, as described in Section 7.1.4.
The reason for this restriction is that the driver cannot account for such global variables in its memory utilization measure. If you need space for large data structures, you can allocate space for them within the heap.
The presence of some undefined behavior is inevitable in malloclab, where we perform unsafe memory operations and generally treat memory as a giant playground. For example, we need to do things like cast between pointer types, perform pointer arithmetic, and write to arbitrary locations in memory. Outside of this lab, it is rarely appropriate to write code in this style, but it is necessary to do so here.
However, we ask you to minimize the amount of undefined behavior by avoiding it when possible. For example, instead of directly casting between pointer types, you should explicitly alias memory through the use of unions. Additionally, you should limit the use of pointer arithmetic to a few short helper functions, as we have tried to do in the handout code.
In the provided baseline code, we use a zero-length array to declare a payload element in the block struct. This is a non-standard compiler extension, but we encourage you to use it for this lab as a way to explicitly declare your block payload.
One important restriction is that a zero-length array must appear as the last element in its containing struct. However, for malloclab, we will also permit the use of a zero-length array as a member of a union, as long as that union is the last element in its containing struct. This will allow you to explicitly alias the payload with other data that you might also want to store in the same memory address. Though this is not portable, we strongly encourage its use compared to the alternative of casting between pointer types.
The use of macro definitions (using #define) in your code is restricted to the following:
– Macros for debugging purposes, whose names must begin with the prefix “dbg ”. See, for example, the debugging macros defined in mm.c. You may create your own, but they must be disabled in submitted code.
•
•
•
8
– Definitions of constants. These definitions must not have any parameters.
The practice of using macros instead of function definitions is now obsolete. Modern compilers (when optimization is enabled) perform inline substitution of small functions, eliminating any inefficiencies due to the use of functions rather than macros. In addition, functions provide better type checking and enable better debugging support.
Here are some examples of allowed and disallowed macro definitions:
#define DEBUG 1 OK #define CHUNKSIZE (1<<12) OK #define WSIZE sizeof(uint64 t) OK #define dbg printf(...) printf( VA ARGS ) OK #define GET(p) (*(unsigned int *)(p)) Not OK #define PACK(size, alloc) ((size)|(alloc)) Not OK
Defines a constant Defines a constant Defines a constant Debugging support Has parameters Has parameters
When you run make, it will run a program that checks for disallowed macro definitions in your code. This checker is overly strict—it cannot determine when a macro definition is embedded in a comment or in some part of the code that has been disabled by conditional-compilation directives. Nonetheless, your code must pass this checker without any warning messages.
• The code shown in the textbook (Section 9.9.12, and available from the CS:APP website) is a useful source of inspiration for the lab, but it does not meet the required coding standards. It does not handle 64-bit allocations, it makes extensive use of macros instead of functions, and it relies heavily on low-level pointer arithmetic. Similarly, the code shown in K&R does not satisfy the coding requirements. You should use the provided code mm.c as your starting point.
• It is okay to look at any high-level descriptions of algorithms found in the textbook or else- where, but it is not acceptable to copy or look at any code of malloc implementations found online or in other sources, except for the allocators described in the textbook, in K&R, or as part of the provided code.
• It is okay to adapt code for useful generic data structures and algorithms (e.g., linked lists, hash tables, search trees, and priority queues) from Wikipedia and other repositories. (This code must not have already been customized as part of a memory allocator.) You must include (as a comment) an attribution of the origins of any borrowed code.
• Your allocator must always return pointers that are aligned to 16-byte boundaries. The driver will check this requirement.
• Your code must compile without warnings.
9
6 Driver Programs
Three driver programs are generated when you run make.
1. mdriver should be used for testing your code’s performance. It checks the correctness, uti-
lization, and throughput of the standard benchmark traces.
2. mdriver-dbg should be used when debugging your program. This is the same program as
mdriver, with two notable differences:
(a) This program is compiled with DEBUG defined, which enables the dbg_ macros at the top of mm.c. Without this defined, functions like dbg_printf and dbg_assert will not have any effect.
(b) Additionally, this program is compiled with optimization level -O0, which allows GDB to display more meaningful debugging information.
3. mdriver-emulate is used to perform additional correctness checks on your program, which can result in additional deductions described in section 7.1.4.
It uses special compilation and memory-management techniques to allow testing your pro- gram with a heap making use of the full, 64-bit address space. In addition to the standard benchmark traces, it will run a set of giant traces with very large allocation requests.
4. mdriver-uninit can be used to detect uses of uninitialized memory, using the MemorySan- itizer tool. It only raises an error when uninitialized memory is actually used, and can track initialization status down to each bit.
6.1 Trace files
The driver programs are controlled by a set of trace files that are included in the traces subdirec- tory. Each trace file contains a sequence of commands that instruct the driver to call your malloc, realloc, and free routines in some sequence. The drivers and the trace files are the same ones we will use when we grade your handin mm.c file.
The traces that count towards both your memory utilization and throughput are marked with a ‘*’ in the output of mdriver. Some of the traces are short traces that do not count toward your memory utilization or throughput but are useful for detecting errors and for debugging, which will not be marked with a ‘*’ next to them.
When the driver programs are run, they will run each trace file multiple times: once to make sure your implementation is correct, once to determine the space utilization, and between 3 and 20 times to determine the throughput.
6.2 Command-line arguments
The drivers accept the following command-line arguments. The normal operation is to run it with no arguments, but you may find it useful to use the arguments during development.
10
• -C: Apply the scoring standards for the checkpoint, rather than for the final submission.
• -f tracefile: Use one particular tracefile instead of the default set of tracefiles for testing
correctness, utilization, and throughput.
• -c tracefile: Run a particular tracefile twice, calling mm init in between each run. Unlike -t, this option tests only for correctness. This option is extremely useful if you want to print out debugging messages.
• -v level: Set the verbosity level to the specified value. Levels 0–2 are supported, with a default level of 1. Raising the verbosity level causes additional diagnostic information to be printed as each trace file is processed. This can be useful during debugging for determining which trace file is causing your malloc package to fail.
• -d level: Controls the amount of validity checking performed by the driver. This is entirely distinct from the DEBUG compile-time define.
At debug level 0, very little checking is done, which is useful when testing performance only.
At debug level 1, the driver checks the payload contents to ensure that they are not overwritten between calls into your code. This is the default.
At debug level 2, the driver will also call into your implementation of mm checkheap. This mode is slow, but it can help identify the exact point at which an error occurs.
Additional arguments can be listed by running mdriver -h.
11
7 Scoring
Malloc Lab is worth 11% of your final grade in the course. This is divided into the Checkpoint and Final submissions, which have separate due dates.
The checkpoint submission, worth 4% of your final grade, is graded as follows:
The final submission, worth 7% of your final grade, is graded as follows:
7.1 Autograded score
driver.pl is the same program that Autolab will use to calculate your score. For the checkpoint submission, it will be run with the -C flag, which will change the grading standards as described later in this section.
The program driver.pl computes the autograded score as follows.
• mdriver is run to obtain the performance index P (Section 7.1.3).
• mdriver-emulate is run, which may result in deductions from this score (Section 7.1.4). Assuming your solution passes all correctness tests, and the graders do not detect any errors in
your source code, two metrics are used to evaluate performance: utilization and throughput.
7.1.1 Utilization
The peak ratio between the aggregate amount of memory used by the driver (i.e., allocated via malloc but not yet freed via free) and the size of the heap used by your allocator. The optimal (but unachievable) ratio equals 100%. You should find good policies to minimize fragmentation in order to make this ratio as high as possible.
The utilization of your allocator U will be calculated as the average utilization across all traces. The associated score will be computed as follows:
• If U ≤ Umin then you will receive no credit for utilization.
• If Umin < U < Umax then your utilization score will scale linearly with U. • If U ≥ Umax then you will receive full credit for utilization.
Autograded score (7.1) Heap checker (7.2)
100 points 10 points
Total
110 points
Autograded score (7.1) Code style (7.3)
100 points 4 points
Total
104 points
12
Version
Umin
Umax
Checkpoint Final
55% 55%
58% 74%
7.1.2 Throughput
Throughput is measured by the average number of operations completed per second, expressed in kilo-operations per second or KOPS. A trace that takes T seconds to perform n operations will have a throughput of n/(1000 · T ) KOPS.
Throughput measurements vary according to the type of CPU running the program. We will compensate for this machine dependency by evaluating the throughput of your implementations relative to those of reference implementations running on the same machine. For information on how this is done, see Appendix A.2.
The throughput of your allocator T will be calculated as the average throughput across all traces. This calculation will be based off the reference throughput Tref which will change from checkpoint to final, and from machine to machine. The associated score will be computed as follows:
• If T ≤ Tmin then you will receive no credit for throughput.
• If Tmin < T < Tmax then your throughput score will scale linearly with T . • If T ≥ Tmax then you will receive full credit for throughput.
The throughput standards are set low enough that we expect your program will significantly exceed the requirements for Tmax. Doing so will greatly reduce the frustration of dealing with the different CPU models and the fluctuations from one run to another caused by the system load.
Note that you will get an indication of your program’s throughput by running it on a Shark machine, but you must submit it to Autolab to get a reliable calibration.
7.1.3 Performance index
Observing that both memory and CPU cycles are expensive system resources, we combine these two measures into a single performance index P , with 0 ≤ P ≤ 100, computed as a weighted sum of the space utilization and throughput:
Version
Tmin
Tmax
Checkpoint Final
0.1 · Tref 0.5 · Tref
0.8 · Tref 0.9 · Tref
Version
Utilization
Throughput
Checkpoint Final
20% 60%
80% 40%
13
That is, in English, for the checkpoint your score will be computed as 20% utilization and 80% throughput, and for the final it will be 60% utilization and 40% throughput.
The mdriver driver program will assign a performance index of 0 if it detects an error in any of the traces.
7.1.4 Autograded deductions
The driver.pl program will also run the mdriver-emulate program (see Section 6), which em- ulates a full 64-bit address space. This may deduct points from your autograded score in the following circumstances:
• If mdriver-emulate fails to run successfully, 30 points will be deducted. This can happen if your code fails to support a full 64-bit address space; for example, if it uses int where a size t is needed.
• If the utilization of a trace differs between mdriver and mdriver-emulate, 30 points will be deducted.
• If your program uses more than 128 bytes of global data (see the Programming Rules in Section 5), then up to 20 points will be deducted.
7.2 Heap Consistency Checker
10 points will be awarded based on the quality of your implementation of mm checkheap. The heap checker will be graded for your checkpoint submission only. It will not be graded in your final submission.
We require that you check all of the invariants of your data structures. Specific items will be dependent on your design, so after making design changes, think about what changes you need to make to your heap checker. Some examples of what your heap checker should check:
• Checking the heap (implicit list, explicit list, segregated list):
– Check for epilogue and prologue blocks.
– Check each block’s address alignment.
– Check blocks lie within heap boundaries.
– Check each block’s header and footer: size (minimum size), previous/next allocate/free bit consistency, header and footer matching each other.
– Check coalescing: no consecutive free blocks in the heap. • Checking the free list (explicit list, segregated list):
– All next/previous pointers are consistent (if A’s next pointer points to B, B’s previous pointer should point to A).
14
– All free list pointers are between mem heap lo() and mem heap high().
– Count free blocks by iterating through every block and traversing free list by pointers
and see if they match.
– All blocks in each list bucket fall within bucket size range (segregated list).
Your heap checker should run silently until it detects some error in the heap. Then, and only then, should it print a message and return false. If it finds no errors, it should return true. It is very important that your heap checker run silently; otherwise, it will produce too much output to be useful on the large traces.
7.3 Style
4 points will be awarded based on the quality of your code style, following the Style Guidelines on the website at http://www.cs.cmu.edu/~213/codeStyle.html. Style will be graded for your final submission only. It will not be graded for your checkpoint submission.
Some points to keep in mind for malloclab in particular:
• Version control. You must commit your code regularly using Git. This allows you to keep track of your changes, revert to older versions of your code, and regularly remind yourself of what you changed and why you made those changes. For specific guidelines on Git usage, see the style guideline.
• Modularity. Your code should be decomposed into functions and use as few global variables as possible. You should use static functions and declared structs and unions to minimize pointer arithmetic and to isolate it to a few places.
• Magic numbers. You should avoid sprinkling your code with numeric constants. Instead, use declarations via #define or static constants. Try, as much as possible, to use C data types, and the operators sizeof and offsetof to define the sizes of various fields and offsets, rather than using fixed numeric values.
• Header comment. Your mm.c file must begin with a header comment that gives an overview of the structure of your free and allocated blocks, the organization of the free list, and how your allocator manipulates the free list.
• Function comments. In addition to the overview header comment, each function must be preceded by a header comment that describes what the function does. Make sure to review the course style guide: we are expecting that for each function, you document at a minimum its purpose, arguments, return value, and any relevant preconditions or postconditions.
• Inline comments. You will want to use inline comments to explain code flow or code that is tricky.
• Extensibility. Your code should be modular, robust, and easily scalable. You should be able to easily change various parameters that define your allocator, without any changes in
15
the actual operation of the program. For example, you should be able to arbitrarily change the number of segregated lists with minimal modifications.
Study the code in mm.c as an example of the desired coding style.
For formatting your code, we require that you use the clang-format tool, which automatically reformats your code according to the .clang-format configuration file. To invoke it, run make format. You are welcome to change the configuration settings to match your desired format. More information is available in the style guideline.
7.4 Handin Instructions
Make sure your code does not print anything during normal operation, and that all debugging macros have been disabled. Ensure that you have committed and pushed the latest version of your code to GitHub.
To submit your code, run make submit or upload your mm.c file to Autolab. Only the last version you submit will be graded.
16
8
•
• • •
•
• •
•
•
Useful Tips
You’ll find debugging macros defined in mm.c that provide contract functions, such as dbg assert. We encourage making liberal use of these contracts to verify invariants and ensure correctness
of your code.
Use the drivers’ -c and -f options to run individual traces. During initial development, using short trace files will simplify debugging and testing.
Use the drivers’ verbose mode. The -V option will also indicate when each trace file is processed, which will help you isolate errors.
Use gdb to help you debug. This will help you isolate and identify out-of-bounds memory references. When debugging, use the mdriver-dbg binary, which is compiled with the -O0 flag to disable optimizations.
Use the Clang static analyzer. The Clang static analyzer is able to find and pinpoint some bugs and provide explanations for them at compile time. To use it, run make clean and then run /usr/local/depot/llvm-7.0/bin/scan-build make and look for errors that it found in your mm.c.
Use gdb’s watch command to find out what changed some value you did not expect to have changed.
Reduce obscure pointer arithmetic through the use of struct’s and union’s. Although your data structures will be implemented in compressed form within the heap, you should strive to make them look as conventional as possible using struct and union declarations to encode the different fields. Examples of this style are shown in the baseline implementation.
Encapsulate your pointer operations in functions. Pointer arithmetic in memory managers is confusing and error-prone because of all the casting that is necessary. You can significantly reduce the complexity by writing functions for your pointer operations. See the code in mm.c for examples. Remember: you are not allowed to define macros—the code in the textbook does not meet our coding standards for this lab.
Your allocator must work for a 64-bit address space. The mdriver-emulate will specifically test this capability. You should allocate a full eight bytes for all of your pointers and size fields. (You can take advantage of the low-order bits for some of these values being zero due to the alignment requirements.) Make sure you do not inadvertently invoke 32-bit arithmetic with an expression such as 1<<32, rather than 1L<<32.
17
9 Office Hours
This lab has a significant implementation portion that is much more involved than the prior labs. Expect to be debugging issues for several hours — there is no way around this.
TAs will NOT:
• Go line-by-line through your program to determine where things go wrong.
• Read your code to determine if everything you are doing is correct. There are simply too many subtle issues to go through everyone’s programs.
TAs will:
• Help you use GDB efficiently to monitor the state of your program.
• Go through conceptual discussions of what you are doing.
Here are some useful things to have ready before a TA comes to you:
• A non-trivial heap checker that exhaustively tests your implementation.
• Documentation of your data structures (drawings are great!)
• A detailed description of problems you are experiencing, and what you’ve looked into so far (not “my malloc doesn’t work and I don’t know why”)
As a reminder, check our office hours schedule on the course website. Start early so you can get help early!
18
10 Strategic Advice
You must design algorithms and data structures for managing free blocks that achieve the right balance of space utilization and speed. This involves a trade-off—it is easy to write a fast allocator by allocating blocks and never freeing them, or a high-utilization allocator that carefully packs blocks as tightly as possible. You must seek to minimize wasted space while also making your program run fast.
As described in the textbook and the lectures, utilization is reduced below 100% due to fragmen- tation, taking two different forms:
• External fragmentation: Unused space between allocated blocks or at the end of the heap
• Internal fragmentation: Space within an allocated block that cannot be used for storing data, because it is required for some of the manager’s data structures (e.g., headers, footers, and free-list pointers), or because extra bytes must be allocated to meet alignment or minimum block size requirements
To reduce external fragmentation, you will want to implement good block placement heuristics. To reduce internal fragmentation, it helps to reduce the storage for your data structures as much as possible.
Maximizing throughput requires making sure your allocator finds a suitable block quickly. This involves constructing more elaborate data structures than is found in the provided code. However, your code need not use any exotic data structures, such as search trees. Our reference implemen- tation only uses singly- and doubly-linked lists.
Here’s a strategy we suggest you follow in meeting the checkpoint and final version requirements:
• Checkpoint. The provided code already meets the required utilization performance2, but it has very low throughput. You can achieve the required throughput by converting to an explicit-list allocator, and then converting that into a segregated free list allocator. Both of these changes will not affect your utilization much.
You will want to experiment with allocation policies. The provided code implements first-fit search. Some allocators attempt best-fit search, but this is difficult to do efficiently. You can find ways to introduce elements of best-fit search into a first-fit allocator, while keeping the amount of search bounded.
Depending on whether you place newly freed blocks at the beginning or the end of a free list, you can implement either a last-in-first-out (LIFO) or a first-in-first-out (FIFO) queue discipline. You should experiment with both.
• Final Version. Building on the checkpoint version, you must greatly increase the utilization and keep a high throughput. You must reduce both external and internal fragmentation. Reducing external fragmentation requires achieving something closer to best-fit allocation.
2 The baseline code meets the required utilization for the checkpoint given a basic implementation of coalesce block. However, you will need to implement coalescing yourself.
19
Reducing internal fragmentation requires reducing data structure overhead. There are mul- tiple ways to do this, each with its own challenges. Possible approaches and their associated challenges include:
– Eliminate footers in allocated blocks. But, you still need to be able to implement coalescing. See the discussion about this optimization on page 852 of the textbook.
– Decrease the minimum block size. But, you must then manage free blocks that are too small to hold the pointers for a doubly linked free list.
– Set up special regions of memory for small, fixed-size blocks. But, you will need to manage these and be able to free a block when given only the starting address of its payload.
See Appendix A.1 for approximate expected quantitative results from these optimizations.
Some advice on how to implement and debug your packages will be covered in the lectures and recitations, as well as the Malloc Lab Bootcamp.
Good luck!
20
A A.1
Performance Evaluation
Approximate Expected Results from Optimizations
Optimization
Utilization
Throughput
Implicit List (Starter Code)
59%
10-100
Explicit Free List
-3
2000-5000
Segregated Free Lists
-
11000
Better Fit Algorithm
59%
Variable
Eliminating Footers in Allocated Blocks
+9%
-
Decreasing Block Size/Mini Blocks
+6%
-20%
Compressing Headers
+2%
-
A.2
Machine Dependencies
You will find that your program gets different throughput values depending on the model of CPU for the machine running the program. Our evaluation code compensates for these differences by comparing the performance of your program to implementations we have written for both the checkpoint and the final versions. It can determine the reference performance for the machine executing the program in two different ways. First, it looks in the file throughputs.txt to see if it has a record for the executing CPU. (Linux machines contain a file /proc/cpuinfo that includes information about the CPU model.) Second, if it does not find this information, it runs reference implementations that are included as part of the provided files.
Throughput information has been generated for the CPUs in the Shark machines, as well as the CPU model used by the Autolab servers, which we refer to as “Autolab C”. The different CPU types used are:
When mdriver runs, it will print out the CPU model ID (a compressed version of the information in /proc/cpuinfo) and the benchmark throughput for that CPU model.
A.3 Performance Points
Observing that both memory and CPU cycles are expensive system resources, we combine these two measures into a single performance index P , with 0 ≤ P ≤ 100, computed as a weighted sum of the space utilization and throughput:
3- indicates no change.
Name
ID
Class Model Clock (GHz)
Shark Autolab C
Intel(R)Xeon(R)CPUE5520@2.27GHz Intel(R)Xeon(R)Gold6132CPU@2.60GHz
Intel Xeon E5520 2.27 Intel Xeon Gold 6132 2.60
21
T−Tmin +(1−w)·Threshold Tmax−T
min
where U is the space utilization (averaged across the traces), and T is the throughput (harmonic mean across the traces). Umax and Tmax are the estimated space utilization and throughput of a well-optimized malloc package, and Umin are Tmin are minimum space utilization and throughput values, below which you will receive 0 points. The weight w defines the relative weighting of utilization versus performance in the score.
U−Umin
P(U,T)=100 w·Threshold Umax−U min
The Threshold function is defined by
Threshold(x) = x,
1 ,
0 ,
x < 0 0≤x≤1 x > 1 .
The values of Tmin and Tmax are computed relative to the performance of reference versions of allocators, with one designed to meet the utilization and throughput goals for the checkpoint, and the other to meet the goals for the final submission. If the reference version provides throughput Tref, then the throughput values used in computing the score are given as:
Tmin = Rmin · Tref Tmax = Rmax · Tref
where the values of Rmin and Rmax differ for the checkpoint and the final versions.
The following table shows the evaluation parameters for the checkpoint and final versions:
The following table summarizes the throughput standards, based on CPU model and checkpoint or final version:
Version
w
Umin Umax
Rmin Rmax
Checkpoint Final
0.2 0.6
0.55 0.58 0.55 0.74
0.1 0.8 0.5 0.9
Machine
Checkpoint
Final
Tmin Tmax Tref
Tmin Tmax Tref
Shark Autolab C
1,211 9,690 12,112 1,319 10,554 13,193
4,004 7,208 8,009 4,630 8,334 9,261
22
B Viewing Heap Contents with GDB
The debugging program gdb can be a valuable tool for tracking down bugs in your memory allocator. We hope by this point in the course that you are familiar with many of the features of gdb. You will want to take full advantage of them.
Unfortunately, the normal gdb commands cannot be used to examine the heap when running mdriver-emulate. In this appendix, we present a brief tutorial on using gdb with your program and describe a helper function that can be used to examine the heap with gdb for both mdriver and mdriver-emulate. In this tutorial, we use the code in mm.c as the reference implementation.
B.1 Viewing the heap without a helper function
A typical gdb session to examine the header of a block on a call to free might go something like this. In the following, all text in italics was typed by the user. The session has been edited to remove some uninteresting parts of the printout.
linux> gdb mdriver
(gdb) break mm free
Breakpoint 1 at 0x4043a1: file mm.c, line 288.
(gdb) run -c traces/syn-array-short.rep
Breakpoint 1, mm free (bp=bp@entry=0x80000eac0) at mm.c:288 (gdb) print bp
$1 = (void *) 0x80000eac0
(gdb) print /x *((unsigned long *) bp – 1)
$2 = 0x41a1
(gdb) quit
A few things about this session are worth noting:
• The function named “free” in mm.c is known to gdb in its unaliased form as “mm free.” You can see that the aliasing is introducted through a macro definition at the beginning of the file. When you use gdb, you refer to the unaliased function names. The unaliased names of other important functions in mm.c include: mm malloc, mm realloc, mm calloc, mem memset, and mem memcpy.
• The gdb command “print /x *((unsigned long *) bp – 1)” first casts the argument to free to be a pointer to an unsigned long. It then decrements this pointer to point to the block header and then prints it in hex format.
• The printed value 0x41a1 indicates that the block is of size 0x41a0 (decimal 16,800), and the lower-order bit is set to indicate that the block is allocated. Looking at the trace file, you will see that the block to be freed has a payload of 16,784 bytes. This required allocating a block of size 16,800 to hold the header, payload, and footer.
23
When we try the same method with mdriver-emulate, things don’t work as well. In this case, we use one of the traces with giant allocations, but the same problem will be encountered with any of the traces.
linux> gdb mdriver-emulate
(gdb) break mm free
Breakpoint 1 at 0x4043b7: file mm.c, line 285.
(gdb) run -c traces/syn-giantarray-short.rep
Breakpoint 1, mm free (bp=bp@entry=0x23368bd380eb2cb0) at mm.c:285 (gdb) print bp
$1 = (void *) 0x23368bd380eb2cb0
(gdb) print /x *((unsigned long *) bp – 1)
Cannot access memory at address 0x23368bd380eb2ca8
(gdb) quit
The problem is that bp is set to 0x23368bd380eb2cb0, or around 2.54×1018, which is well beyond the range of virtual addresses supported by the machine. The mdriver-emulate program uses sparse memory techniques to provide the illusion of a full, 64-bit address space, and so it supports very large pointer values. However, the actual addressing has an overall limit of 100 MB of actual memory usage.
B.2 Viewing the heap with the hprobe helper function
To support the use of gdb in debugging both the normal and the emulated version of the memory,
we have created a function with the following prototype:
void hprobe(void *ptr, int offset, size t count);
This function will print the count bytes that start at the address given by summing ptr and offset. Having a separate offset argument eliminates the need for doing pointer arithmetic in your queury.
Here’s an example of using hprobe with mdriver:
linux> gdb mdriver
(gdb) break mm free
Breakpoint 1 at 0x4043a1: file mm.c, line 288.
(gdb) run -c traces/syn-array-short.rep
Breakpoint 1, mm free (bp=bp@entry=0x80000eac0) at mm.c:288 (gdb) print bp
$1 = (void *) 0x80000eac0
(gdb) print hprobe(bp, -8, 8)
Bytes 0x80000eabf…0x80000eab8: 0x00000000000041a1
(gdb) quit
24
Observe that hprobe is called with the argument to free as the pointer, an offset of −8 and a count of 8. The function prints the bytes with the most significant byte on the left, just as for normal printing of numeric values. The range of addresses is shown as HighAddr…LowAddr. We can see that the printed value is identical to what was printed using pointer arithmetic, but with leading zeros added.
The same command sequence works for mdriver-emulate:
linux> gdb mdriver-emulate
(gdb) break mm free
Breakpoint 1 at 0x4043b7: file mm.c, line 285.
(gdb) run -c traces/syn-giantarray-short.rep
Breakpoint 1, mm free (bp=bp@entry=0x23368bd380eb2cb0) at mm.c:285 (gdb) print bp
$1 = (void *) 0x23368bd380eb2cb0
(gdb) print hprobe(bp, -8, 8)
Bytes 0x23368bd380eb2caf…0x23368bd380eb2ca8: 0x00eb55b00c8f1ed1 (gdb) quit
The contents of the header indicate a block size of 0xeb55b00c8f1ed0 (decimal 66,240,834,140,315,344), with the low-order bit set to indicate that the block is allocated. Looking at the trace, we see that the block to be freed has a payload of 66,240,834,160,315,328 bytes. Sixteen additional bytes were required for the header and the footer.
Part of being a productive programmer is to make use of the tools available. Many novice pro- grammers fill their code with print statements. While that can be a useful approach, it is often more efficient to use debuggers, such as gdb. With the hprobe helper function, you can use gdb on both versions of the driver program.
25