Performance Analysis of BSTs in System Software

Ben Pfa Stanford University Department of Computer Science

blp@cs.stanford.edu

Abstract

agement and networking, and the third analyzes a part of a source code cross-referencing tool. In each Binary search tree (BST) based data structures, such case, some test workloads are drawn from real-world as AVL trees, red-black trees, and splay trees, are of- situations, and some reect worst- and best-case inten used in system software, such as operating system put order for BSTs. kernels. Choosing the right kind of tree can impact performance signicantly, but the literature oers few empirical studies for guidance. We compare 20 BST We compare four variants on the BST data strucvariants using three experiments in real-world scenar- ture: unbalanced BSTs, AVL trees, red-black trees, ios with real and articial workloads. The results in- and splay trees. The results show that each should be dicate that when input is expected to be randomly or- preferred in a dierent situation. Unbalanced BSTs dered with occasional runs of sorted order, red-black are best when randomly ordered input can be relied trees are preferred; when insertions often occur in upon; if random ordering is the norm but occasional sorted order, AVL trees excel for later random access, runs of sorted order are expected, then red-black trees whereas splay trees perform best for later sequential should be chosen. On the other hand, if insertions or clustered access. For node representations, use of often occur in a sorted order, AVL trees excel when parent pointers is shown to be the fastest choice, with later accesses tend to be random, and splay trees perthreaded nodes a close second choice that saves mem- form best when later accesses are sequential or clusory; nodes without parent pointers or threads suer tered. when traversal and modication are combined; maintaining a in-order doubly linked list is advantageous For each BST data structure variant, we compare when traversal is very common; and right-threaded ve dierent node representations: plain, with parnodes perform poorly. ent pointers, threaded, right-threaded, and with an in-order linked list. Combining traversal and modication in a BST with plain nodes requires extra 1 Introduction searches, although there are fewer elds to update. OS kernels and other system software often use bi- Parent pointers are generally fastest, as long as the nary search tree (BST) based data structures such cost of an additional pointer eld per node is not as AVL trees, red-black trees, or splay trees. Choos- important. Threaded representations are almost as ing the right tree and node representation can impact fast as parent pointers, but require less space. Rightthe performance of code that uses these data struc- threaded representations fare poorly in all of our extures. Surprisingly, there has been little empirical periments, as do in-order linked lists. study of the relationship between the algorithms used for managing BST-based data structures and perforThe remainder of this paper is organized as follows. mance characteristics in real systems [1, 2, 3]. This Section 2 describes the data structures and node reppaper attempts to ll this gap by thoroughly analyzresentations being compared and section 3 states our ing the performance of 20 dierent variants of binary hypotheses for their performance. Section 4 describes search trees under real and articial workloads. the experimental platform used for comparisons. SecWe present three experiments. The rst two exam- tion 5 presents the experiments and discusses their ine data structures used in kernels in memory man- results individually, while section 6 discusses the re An extended abstract of this paper appeared in the proceedings of SIGMETRICS/Performance 2004.

sults as a whole. Section 7 covers related work, and section 8 draws overall conclusions.

BST Variants
2 1
(a)

3 6 4 5 3 2 1
(b) Figure 1: (a) A threaded BST, with threads drawn as dotted lines. (b) A right-threaded BST. requires a search starting from the root. Figure 1(b) shows a right-threaded BST.

This paper focuses on two dimensions within the BST design space. The rst dimension under consideration is the choice of data structure; the second, choice of BST node representation.

8 7 9

2.1

Data Structure

In an ordinary, unbalanced BST, insertion of data items in a pathological order, such as sorted order, causes BST performance to drop from O(lg n) to O(n) per operation in a n-node tree [4, 5, 6]. One solution is a balanced tree, which uses balancing rules to bound tree height to O(lg n). The most popular balanced trees are AVL trees and red-black trees, which limit the height of an n-node tree to no more than 1.4405 lg(n + 2) 0.3277 and 2 lg(n + 1), respectively [7, 8, 9, 10]. Another solution is a self-adjusting tree, that is, one that dynamically moves frequently accessed nodes near the trees root. The most popular selfadjusting tree structure is the splay tree, which rotates or splays each node to the root at time of access [11]. Splay tree amortized performance has been shown to be optimal in several important respects [11, 12].

6 4 5 7 8 9

Performance Hypotheses

Before describing the experiments, we will take a brief look at the qualitative performance we expect from operations on each tree variant.

2.2

Node Representation 3.1 Data Structure

When data items are inserted in random order, any BST-based data structure should provide acceptable behavior. There is no need for complex rebalancing algorithms, because ordinary BSTs will produce acceptably balanced trees with high likelihood. Thus, for random insertions, data structures with the least extra overhead (above ordinary BSTs) should yield the best performance. The red-black balancing rule is more permissive than the AVL balancing rule, resulting in less superuous rebalancing, so red-black trees should perform better for random insertions. We suspect that splay trees will be slower than either AVL or red-black trees on random insertions because they splay the inserted node to the root on every insertion, whereas the other balanced trees only perform balancing as necessary. On the other hand, when data items are inserted in sorted order, ordinary BST behavior should be the slowest by far. In such a case, the stricter the balancing rule, the shorter in height the tree that should be produced. We therefore expect to see that for pathological input, AVL trees should perform better than 2

At a minimum, a BST node contains a data item and left and right child pointers. Ecient traversal of such a ordinary BST requires maintenance of a stack of nodes to revisit. Most kinds of tree modication during traversal also require that the stack be regenerated with an O(lg n)-cost search operation. One obvious way to solve these problems is to add a parent pointer to the node, yielding a tree with parent pointers. Another simple technique is to add a predecessor and a successor pointer to each node, maintaining a doubly linked list of in-order nodes; we will call this a linked list tree for short. Alternatively, each right child pointer that would otherwise be null can be used to point to the nodes in-order successor, and similarly for left child pointers and node predecessors. These successor and predecessor pointers are called threads, and a tree with threads is a threaded tree [13]. Figure 1(a) shows a threaded BST, with threads drawn as dotted lines. When only successors are of interest, not predecessors, it is possible to thread right child pointers only, producing a right-threaded tree in which nding the successor of a node is fast, but nding a predecessor

red-black trees. Known linear performance for sequential access patterns in splay trees [12] suggests that splay trees should perform well, but it is dicult to predict how they will do relative to AVL and red-black trees. Some data structures require extra memory on top of that required by ordinary BSTs: AVL trees require 2 extra bits per node and red-black trees require 1 extra bit. Due to padding, these bits actually add 4 bytes to each node in libavl. They could be absorbed into other elds, but not in portable C code and at a performance cost. Splay trees do not require extra memory.

branch, similar to following links. In right-threaded trees, each rotation requires a two-way branch and sometimes 1 extra assignment. Node representations also have dierent memory requirements. Parent pointers add a single pointer eld to each node; linked list nodes add two. Threaded nodes add 2 tag bits to each node used to distinguish threads from child pointers, and rightthreaded similarly nodes add 1 tag bit. In AVL and red-black trees, tag bits can be absorbed into otherwise unused padding at no extra cost.

3.2

Node Representation

Experimental Platform

To distinguish threads from child pointers, threaded nodes require an extra check for each link followed, costing extra time. This penalty could be signicant because following links is such a common operation in examining and manipulating binary trees. Following a link from a right-threaded node is the same as plain nodes on the left side, and the same as threaded nodes on the right side. Ideally, traversal operations, for nding in-order successors and predecessors, should achieve O(1) performance. The actual implementation and performance of these functions is a key dierence between node representations. Based on the complexity of operation in the normal case, the table below lists them in suspected order of fastest to slowest: linked list O(1) traversal by a single branchless pointer dereference. threads O(1) traversal by following a single pointer upward or a series downward. right threads O(1) successor operation, as for threaded nodes, but nding a nodes predecessor requires an O(lg n) search from the tree root when it starts from a node with no left child. parent pointers O(1) traversal by following O(1) pointers upward or downward in the tree. plain O(1) traversal by maintaining an explicit stack. Performance degrades to O(lg n) if the tree is modied. Rotation operations, used to maintain balance in balanced trees, suer overhead under many node representations. With parent pointers, each rotation requires 3 extra pointer assignments, doubling the number of assignments. In threaded trees, each rotation requires 1 or 2 extra assignments and a two-way 3

If we consider ordinary BSTs, AVL trees, red-black trees, and splay trees, for each of the ve kinds of tree nodes described above, there are 20 dierent ways to implement binary search tree-based data structures. To test 12 of the 20 variations, we used GNU libavl 2.0.1, a free software library in portable ANSI/ISO C [14]. GNU libavl does not currently include splay trees or linked list node representations, so the 8 remaining variants were implemented for this paper. All of the libavl tree implementations implement the same interface, summarized in Table 1. The interface divides into two categories of functionality: set functions and traverser functions. The set interface manipulates trees in terms of a set abstraction, that is, as an unordered collection of unique items.1 . The probe function combines search and insertion and the insert function inserts a new item into a tree when it is known not to contain an item with an equal key. Function delete removes an item. The find function searches for an item whose key equals a given target. The traverser interface allows a program to iterate through the items in a tree in sorted order. Each traverser tracks the position of a node in the tree, with the guarantee that a traverser remains valid across any series of tree modications as long as its own node is not deleted. A traverser may also have a null current position. Functions that create traversers include t first, to create a traverser with the minimum value in the tree as its current position; t last, for the maximum value; t find, for a value equal to a target; and t insert, to insert a new item and create a traverser at its position. Operations on traversers include t next and t prev, which move a traverser to the next or previous node, respectively, in sorted order and return its data.
1 libavl does not directly support trees that contain duplicates

find(x ): searches for an item that equals x insert(x ): inserts x (for use when x is known not to be in the tree) probe(x ): combines search and insertion: searches for x and inserts it if not found delete(x ): deletes an item equal to x if one exists in the tree t first(t): positions traverser t at the minimum item in the tree t last(t): positions traverser t at the maximum item in the tree t find(t, x ): positions traverser t at an item equal to x or at the null position if none exists t insert(t, x ): positions traverser t at an item equal to x, inserting x if none exists t next(t): advances traverser t to the next item in in-order t prev(t): advances traverser t to the previous item in in-order Table 1: Summary of important libavl functions.

Experiments

This section discusses three experiments carried out to test data structure performance. Each experiment takes advantage of a dierent property of BSTs. Two of the three experiments are simulations of actual uses of balanced binary trees within the Linux kernel, version 2.4.20. All three are intended to represent substantial real-world uses for BST-based structures. The primary platform for the experiment, for which times are reported, was a Compaq Armada M700 with 576 MB RAM and a Mobile Pentium III processor at 500 MHz with 256 kB cache, running Debian GNU/Linux unstable with the Linux 2.4.22 kernel. The compiler used was GCC 2.95.4 with ags -O3 -DNDEBUG. The experiments were also run on other x86-based machines, including Pentium III and Pentium IV desktop machines, and under other versions of GCC, including 3.2.2 and 3.3. The dierences between processors and compiler versions were minor. The one signicant dierence that did appear is described as part of its experimental results, in section 5.2.1.

binaries also have code and data VMAs for each dynamic library. Processes can create an arbitrary number of VMAs, up to the limit imposed by the operating system or machine architecture, by mapping disk les into memory with the mmap system call. A complementary system call, munmap, can partially or entirely eliminate existing VMAs [15]. Since VMAs vary in size over several orders of magnitude, from 4 kB to over 1 GB, conventional hash tables cannot keep track of them eciently. Hash tables also do not eciently support range queries needed to determine which existing VMAs overlap the range requested by a mmap or munmap call. BST-based data structures, on the other hand, have both these properties, so many kernels use BSTs for keeping track of VMAs: Linux before 2.4.10 used AVL trees, OpenBSD and later versions of Linux use red-black trees, FreeBSD uses splay trees, and so does Windows NT for its VMA equivalents [16]. When BSTs store intervals that can overlap, efcient performance requires the tree structure to be augmented with an extra eld, as in interval trees [17]. VMAs do not overlap, so this is unnecessary as long as care is taken during insertion operations. Interval trees are not used here, nor are they used by any of the kernels mentioned above. The rst experiment simulates VMA activity during program execution. Modications to VMA tables by mmap and munmap calls are simulated by tree modications. Test sequences were drawn from four sources, primarily by instrumenting the behavior of real programs: Mozilla 1.0 [18] over the course of a brief browsing session. In total, mmap is called 1,459 times and munmap 1,264 times. At the peak there are 108 VMAs, with an average of 98. VMware GSX Server 2.0.1 x 86 virtual machine monitor [19] over the boot sequence and brief usage of a Debian GNU/Linux virtual machine. VMware GSX Server divides itself into several processes; behavior of only the most VMAintensive of these is simulated. In total, mmap and munmap are each called 2,233 times. At peak there are 2,233 VMAs, with an average of 1,117. Squid web cache 2.4.STABLE4 running under User-Mode Linux 2.4.18.48, which relies on mmap for simulating virtual memory. In total, mmap is called 1,278 times and munmap 1,403 times. There are at most 735 VMAs at any time, 400 on average. A synthetic test set consisting of 1,024 mmap calls followed by 1,024 munmap calls, all of which begin 4

5.1

Virtual Memory Areas

Each process in a Unix-like kernel has a number of virtual memory areas (VMAs). At a minimum, a statically linked binary has one VMA for each of its code, data, and stack segments. Dynamically linked

Address Space

Address Space

Address Space

544 1088 1632 2176 2720

892 1784 2676 3568 4460

528 1056 1584 2112 2640

Address Space
0

409 818 1227 1636 2045

(a)

Operations

(b)

Operations

(c)

Operations

(d)

Operations

Figure 2: Call sequences in (a) Mozilla 1.0, (b) VMware GSX Server 2.0.1, (c) squid running under UserMode Linux 2.4.18.48, and (d) random test sets. Part (b) omits one mmap-munmap pair for memory region 0x20000000 to 0x30000000 and (c) omits address space gaps; the others are complete. at random page-aligned locations and continue for a random number of pages. This is unrealistic, provided as an example of the best case for an ordinary BST. Figure 2 depicts the sequence of calls in each test set. Notably, the VMware GSX Server test set is close to being a worst case for unbalanced BSTs because of its sequence of one-page mmaps at sequential virtual addresses. The Linux kernel (and the simulation) never merges VMAs that refer to les, even when they map sequential blocks within the same le, so each of these calls creates a new VMA. On the other hand, the random test set is the best case for unbalanced BSTs. The function called most often in the VMA simulation, 20,110 times among the four test sets, is t equal range, a function implemented for this simulation that nds the range of tree nodes that overlap a given range as a pair of traversers. Other functions called often are t next, 11,051 times; insert, 5,994 times; delete, 5,044 times; and t prev, 4,460 times. Use of other tree functions is insignicant. Table 2 shows the time, in seconds, to run 1,000 iterations of each test set using each of the 20 tree types, and the number of comparisons performed in a single run of each. 5.1.1 Real-World Data Sets comparison count numbers: in each case, the next larger comparison count was 2.0 to 3.4 times that of the splay tree. The tendency of the VMware and Squid test sets toward sequential access is also a factor, since such access patterns have been proved to take linear time in splay trees [12]. As expected, ordinary BSTs are slower than any other data structure for the real-world tests, by more than an order of magnitude in the pathological VMware test set. The Mozilla test with plain node representation, where ordinary BSTs are the fastest of the four, is exceptional because of a detail of the libavl implementation, in that it uses a stack of xed maximum size for in-order traversal in plain BSTs.2 When traversal encounters a tree branch deep enough to overow the stack, it rebalances the entire tree, reducing the trees height to the minimum possible for its number of nodes. Insertion, deletion, and search operations on the resulting tree are then, initially, guaranteed to run in O(lg n) time. In each run of the Mozilla and Squid test sets such forced rebalancing occurs three times for plain BSTs, and once in the VMware test set. (Plain splay trees have similar xed-sized stacks for traversal, but the test does not trigger it.) In these data sets, the AVL tree implementations were consistently faster than red-black trees, by up to 20%. Table 2 shows that there is a corresponding 12% to 32% increase in the number of comparisons from AVL to red-black, which suggests that the stricter AVL balancing rule may be responsible. Further investigation shows that although the average internal path length for the red-black trees is only about 3% longer than for the AVL trees, the path of maximum length is 16% to 29% longer in both test sets, as shown in table 3. The conclusion is that although both AVL and red-black trees globally balance the trees about as well, the AVL balancing rule
2 The

The Mozilla, VMware, and Squid data sets are considered real world data sets because they are drawn directly from real programs. Splay trees shine in the three real-world tests. They beat all of the other tree types by a wide margin, bettering the best of the competition by 23% to 40% each time. The reason lies in the test sets high locality of reference, as shown in Figure 2, along with the splay trees ability to keep frequently used nodes near the top of the tree. This is demonstrated by the 5

maximum size is 32 entries by default.

test set Mozilla

VMware

Squid

random

representation plain parents threads right threads linked list plain parents threads right threads linked list plain parents threads right threads linked list plain parents threads right threads linked list

time (seconds) BST AVL RB 4.49 4.81 5.32 15.67 3.65 3.78 16.77 3.93 3.95 16.91 4.07 4.20 16.31 3.64 4.35 208.00* 8.72 10.59 447.40* 6.31 7.32 445.80* 6.91 8.51 446.40* 6.88 8.59 472.00* 7.35 8.60 7.34 4.41 4.67 12.52 3.69 3.80 13.44 3.92 4.18 14.46 4.17 4.27 13.13 4.02 4.19 2.83 2.81 2.86 1.63 1.67 1.64 1.64 1.74 1.68 1.92 1.96 1.93 1.46 1.54 1.51

splay 2.71 2.63 2.67 2.68 2.74 3.77 3.62 3.64 3.51 3.45 2.84 2.64 2.70 2.86 2.65 3.43 1.94 2.02 2.22 1.74

BST 71,864 635,957 635,957 634,844 559,062 7,503,740 14,865,389 14,865,389 14,872,042 14,865,389 250,883 487,818 487,818 511,903 487,818 37,396 28,123 28,123 30,598 27,845

comparison AVL 62,523 54,842 54,842 53,962 46,904 137,081 122,087 122,087 122,076 122,087 67,079 62,467 62,467 63,189 62,467 34,090 25,958 25,958 28,438 25,808

count RB 79,481 68,942 68,942 69,357 69,232 193,984 175,861 175,861 175,886 175,861 77,836 72,867 72,867 71,629 72,867 34,176 25,983 25,983 28,393 25,605

splay 23,398 22,284 22,284 22,241 22,240 32,190 35,733 29,758 22,725 29,831 30,086 29,338 28,156 30,199 28,173 44,438 33,036 33,036 34,672 31,965

Table 2: Time, in seconds, for 1,000 runs of each VMA simulation test set, and number of comparisons during each run. *Estimated based on 10 runs. leads to better local balancing in the important places for test sets like these with strong locality. The splay tree results for maximum path length are deceptive because a splay operation on a node roughly halves the path length to every node along its access path, so that only nodes accessed infrequently will lie along the path of maximum length. Considering AVL and red-black tree implementations for the real-world test sets, in all cases plain representations were the slowest and parent pointer representations were the fastest. Given the experimental operation mix, there are two causes for this behavior. First, the plain representation is slow because insertions and deletions invalidate the stacks required for traversal with plain nodes, forcing t next to perform an extra search; the right-threaded representation is slow because of the use of t prev. Second, parent pointers are faster than threaded representations because distinguishing threads from ordinary child pointers requires an extra step. 5.1.2 Random Data Set ordinary BST, red-black tree, AVL tree, splay tree. Intuitively, the fastest implementations should be the ones that perform the least extra work, because random order is ideal for an ordinary BST. To verify this, we must have some means to measure extra work; because the tested balanced trees all rebalance in terms of rotations, counting rotations is one reasonable way. For the random data set, ordinary BSTs make no rotations, red-black trees make 209, AVL trees make 267, and splay trees make 2,678.3 This ordering neatly coincides with observed performance tendencies, further explaining why splay trees are consistently separated from the other tree types by a relatively wide gap. Along the other dimension, speed tended to be arranged from fastest to slowest in the order of linked list nodes, nodes with parent pointers, threaded nodes, right-threaded nodes, and plain nodes, indicating that performance favors representations that make in-order traversal speedy. Splay trees consistently performed worst, and required the most comparisons, within each node representation category in the random test. This reects
3 Number of rotations is generally independent of node representation. Numbers for right-threaded trees do dier slightly due to a specialized deletion algorithm.

For the random data set, all of the implementations were clustered within a relatively small range of about 2.3 times variation. Within each node representation, speed tended to be ordered from fastest to slowest as 6

test set Mozilla VMware Squid

AVL 468 : 6.8 9,471 : 9.2 2,903 : 8.5

red-black 485 : 8.1 9,723 : 13.1 3,000 : 10.7

splay 1,227 : 25.6 426,378 : 572.3 20,196 : 100.2

Table 3: Average internal path length and average maximum path length, respectively, for each kind of tree in the real-world test sets.

test set representation Mozilla parents threads VMware parents threads Squid parents threads random parents threads

BST AVL 11.12 3.25 11.91 3.52 331.00* 5.42 325.00* 5.96 11.22 3.54 12.10 3.78 1.60 1.64 1.61 1.71

RB splay 3.17 2.53 3.30 2.67 6.35 3.22 7.42 3.55 3.62 2.53 3.99 2.70 1.62 1.90 1.65 2.02

the work required to splay each node accessed to the root of the tree, in the expectation that it would soon Table 4: Time, in seconds, for 1,000 runs of VMA be accessed again. This eort is wasted for the ran- simulation test sets optimized with t delete. *Estimated based on 10 runs. dom set because of its lack of locality. 5.1.3 Optimization a comment in the source notes: [21] Such an implementation has been chosen not just for fun. Its a way to prevent easy and ecient DoS attacks by creating hash collisions. A huge amount of long living nodes in a single hash slot would signicantly delay lookups. . . Techniques such as universal hashing [22] can make hash collision attacks signicantly more dicult. On the other hand, balanced tree techniques make them impossible, so they are a way to dodge an entire class of potential problems. The peer AVL tree is used as a second-level cache only. The rst time a packet is sent to a given host, a node with its IP address is inserted into the tree. A direct pointer to this node is also put into the route cache, which is the rst-level cache. As long as the hosts route remains cached, the AVL tree is not consulted again for that host. If the route is dropped from the cache (generally due to inactivity), and later added back in, the AVL tree is searched a second time. If, on the other hand, the route is dropped and remains inactive, sometime later the hosts peer cache AVL tree entry will be dropped. Peer cache entries not recently used can also be dropped if the cache grows too much, but entries are never dropped while in use in the route cache. The simulation attempts to reproduce the most important features of the peer cache. The route cache is modeled heuristically, not directly. All other features described above are directly reproduced. Two scenarios are modeled: Web server under normal load. A new host with a randomly selected IP address makes its rst request every 10 ms to 1.5 s, with a mean of
arate identication counters for TCP and UDP protocols.

One feature missing from libavl is the ability to directly delete the node referenced by a traverser. Such an operation can avoid a search from the root to the node to delete and thereby reduces deletion time to O(1). For the purposes of this paper such a t delete function was written for implementations with threads and parent pointers. Plain representations are unlikely to benet because most changes invalidate all traverser stacks, and implementation for right-threaded representations appears to always require a search, so t delete was not implemented for those representations. Table 4 shows the timings when t delete is substituted for delete where appropriate. Up to 16% improvement was observed for the balanced tree implementations, even larger for the unbalanced trees. Based on these results, we suggest that future versions of libavl should include a t delete function. A related function insert before, for O(1) insertion with a known insertion point, was also implemented and found to give this experiment no noticeable performance benet, so detailed analysis is omitted.

5.2

Internet Peer Cache

RFC 791 [20] requires that each IP packet sent by an internet host be stamped with a 16-bit identication eld that must be unique for that source-destination pair and protocol for the time the datagram will be active in the internet system. Most TCP/IP implementations ensure this by maintaining a counter for each host they contact, incrementing a hosts counter once for every packet transmitted to it. The Linux kernel keeps track of identication data using an AVL tree indexed by peer IP address.4 As
4 Although

RFC 791 allows it, Linux does not maintain sep-

test set normal

attack

representation plain parents threads right threads linked list plain parents threads right threads linked list

BST 4.74 3.94 3.99 5.52 4.93 * * * * *

AVL 5.10 4.07 4.45 5.71 5.40 3.76 2.65 2.97 4.04 4.10

RB splay 5.06 7.70 3.78 7.19 4.17 13.25 5.64 8.29 5.25 8.41 4.32 4.14 2.92 3.31 3.37 7.77 4.77 4.62 4.46 4.78

Table 5: Times, in seconds, for 500 and 5 runs, respectively, of the normal and attack scenario simulations. *Greater than 60 seconds for a single run. 38 ms. Each currently connected host makes a request every 10 ms to 10 min, with a mean of 30 s, and hosts make a mean of 9 requests each. Attack. A new host with a IP address consecutively greater than the previous makes a request every 10 ms for 1,000 s. Hosts otherwise behave as above, in order to ensure that they stay cached for some time. In each scenario, new peers continue to arrive for 1,000 s simulated time, and then the simulation continues with no new peers until the cache becomes empty. As a result, insertion (insert or t insert) and deletion (delete or t delete) functions are called the same number of times for each test set, 2,691 times each for the normal web server test set and 137,146 times for the attack test set. Due to the read-write locks used by the kernel implementation, a find operation is performed before every insertion. Along with the searches performed to look up peer cache entries expired from the route cache, this makes find the most-used function at 3,884 calls in the normal test set and 147,942 in the attack test set. Use of other libavl functions is unimportant. Table 5 shows the time required for multiple simulated runs of each test set using each tree type. Unbalanced binary trees were not simulated in the attack scenario because of pathological behavior of the algorithm in that case. Ordinary BSTs and splay trees are included in the table, but would not be suitable choices in this scenario because they do not provide guaranteed performance. 5.2.1 Normal Data Set

Within the normal data set, if we sort the representations by speed within each type of tree, in three 8

of four cases the order is the same. From fastest to slowest, this order is: parent pointers, threads, plain, linked list, and right threads. The speed of parent pointer and thread representations reects the ability to use t delete for O(1) deletion. Among the other representations, plain is fastest because it is not hampered by manipulation of a linked list or right threads, which do not provide any benets for this experiment. Splay trees are the exception, for which the threaded representation is by far the slowest. After some investigation, the problem turned out to be an unexpected artifact of the libavl implementation of threaded splay trees. In particular, the routine that implements splaying moves up the splay tree to the root using a well-known algorithm for nding the parent of a threaded node [13, exercise 19]. This algorithm is O(1) on average for randomly selected nodes in a threaded tree, and its wide use in libavls threaded tree implementations shows that it has reasonable performance in most cases. Unfortunately, its pattern of use for splaying coupled with the typical structure of a splay tree leads to poor performance. (The same eect is visible, to a lesser degree, in the threaded splay tree performance shown earlier in Table 2.) To x this performance anomaly, the implementation of threaded splay trees was modied to maintain a stack of parent pointers for splaying purposes, instead of depending on the algorithm for nding a nodes parent. The revised code brought the threaded splay tree runtime down to 8.11 s for the normal case and 4.47 s for the attack case, which are much more competitive times. The revised performance is still not entirely consonant with the other threaded trees, probably due to the need to maintain the stack. (Similar improvement was found when the revised code was substituted in the rst experiment.) When this experiment was run on a Pentium IV system, instead of the Pentium III-based primary experimental platform, the results changed signicantly for splay trees. In particular, the ordering from fastest to slowest changed to parent pointers, linked list, threads, right threads, and plain. We speculate that the Pentium IV has extra-large penalties on branches, favoring parent pointer and linked list representations which require fewer branches than threaded or right-threaded nodes. In the normal case, within each node representation, speed is consistently, from fastest to slowest, in the order of unbalanced trees, red-black trees, AVL trees, and splay trees. This matches our hypotheses for performance of insertion in random order (see 3). Red-black trees with parent pointers are an unex-

plained anomaly. 5.2.2 Attack Data Set

In the attack case, AVL trees perform better than red-black trees because the AVL balancing rule does a better job of keeping the tree height to a minimum at the important point, the point of the next insertion: the maximum path length in the AVL tree, averaged over all operations, is 14.6, whereas in the red-black tree it is 21.5. On the other hand, the average internal path length across all the operations varies by less than 1% between AVL and red-black trees. Splay trees fare poorly for the attack data set, which is somewhat surprising given that sequential accesses in splay trees perform in linear time. Examination of the data set revealed that although peer IP addresses appear in sorted order, the subsequent randomness in request spacing is sucient to create a great deal of disorder: over the 285,088 requested handled by the data set, there is an average distance of 22,652 operations between requests with sequential IP addresses.

5.3

Cross-Reference Collator

For the normal case, splay trees are consistent winners, taking advantage of locality, resulting from the tendency of identiers to recur within a few lines of rst mention. Red-black trees are consistently slightly faster than AVL trees, reecting the more relaxed red-black balancing rule that results in less rebalancing. Finally, the times for unbalanced trees are find . | extract-ids | sort | merge-ids only at most 11% greater than for AVL trees, demonstrating that the natural identier ordering tends towhere extract-ids reads a list of les from its input, ward randomness. extracts their interesting identiers, and writes them along with lename and line number to its output, 5.3.2 Sorted Data Set
5 Implementations based on skip lists, etc., are also possible. The GNU/HP/SGI implementation uses red-black trees with parent pointers.

The two previous experiments focused on problems that were solved especially well with the use of BSTbased data structures. The third experiment is somewhat dierent in that its problem can be better solved with other techniques; in particular, by use of a hash table during accumulation of data followed by a nal sorting pass. It remains interesting for at least one reason: the C++98 standard library [23] contains set and map template classes based on balanced trees,5 but none based on hashes. For that reason, C++ programmers interested in convenience and portability may select one of these standard templates where a hash-based implementation would be more appropriate. Cross-references of identier usage can be useful in software development, especially when source code used as a reference is available as hard copy only. One way to construct such a tool is to divide it into parts connected through a Unix pipeline, like so:

merge-ids collects adjacent lines for identical identiers into a more readable format, and find and sort are the standard Unix utilities. This experiment examines the performance of a utility to take the place of the nal two programs, implemented using libavl. This program, here called a cross-reference collator, inserts a set of identiers into a single tree as they are read. In addition, each identier has its own tree that contains names of the les in which the identier appears. Furthermore, each le has a set of line numbers attached to it corresponding to the lines in the le on which the identier appears. After all of the input has been read, the collator traverses the data structure to produce a fully sorted cross-reference listing. The input used for testing the collator was contents of directory include/linux within the source code for Linux 2.4.18. When extract-id is used, this directory contains 118,639 interesting identiers, or 47,952 dropping duplicates, in 841 les. With this test set, the collator utility creates a total of 47,953 trees (one for each unique identier plus one to organize the identiers) and makes 237,278 calls to probe (two for each identier). It also calls t next 159,815 times while composing output. Table 6 shows the times, in seconds, for 5 runs of the cross-reference collator over the test set for each of libavls table implementations. The test set labeled normal is the normal case where the collators input is presented in the order it is produced by extract-ids. Sets sorted and shued present the articial cases where its input is arranged in order by identier or in random order, respectively, beforehand. Splay trees are tested, but their amortized performance guarantees do not meet the per-operation requirements of the C++98 standard. Unbalanced BSTs also do not meet these requirements. 5.3.1 Normal Data Set

As usual, splay trees are the big winners in the sorted case, beating all other implementations by 28% or 9

test set normal

sorted

shued

representation plain parents threads right threads linked list plain parents threads right threads linked list plain parents threads right threads linked list

BST 5.22 4.97 5.00 5.12 5.05 * * * * * 5.98 5.80 5.80 5.89 5.88

AVL 4.62 4.47 4.63 4.66 4.79 4.21 4.04 4.25 4.24 4.31 5.80 5.68 5.77 5.80 6.06

RB splay 4.49 4.03 4.33 4.00 4.51 4.03 4.57 4.06 4.59 4.03 4.90 2.91 4.70 2.90 5.02 2.89 5.02 2.87 4.98 2.83 5.69 6.54 5.51 6.61 5.65 6.56 5.71 6.63 5.84 6.64

test set normal

shued

representation plain parents threads right threads linked list plain parents threads right threads linked list

BST 3.24 3.10 3.12 3.21 3.19 3.23 3.13 3.15 3.22 3.24

AVL 3.09 2.94 3.04 3.06 3.21 3.28 3.12 3.20 3.22 3.43

RB splay 3.01 2.80 2.86 2.81 2.98 2.82 3.02 2.85 3.08 2.87 3.24 3.51 3.05 3.55 3.15 3.55 3.19 3.60 3.32 3.64

Table 7: Times, in seconds, for 50 runs of a reduced version of the cross-reference collator for each kind of tree designed to t within the processor cache.

Table 6: Times, in seconds, for 5 runs of the unsorted P6 performance counters [24] directly conrms that and sorted versions of the cross-reference collator for reducing the test set size reduces additional cache each kind of tree. *Pathological case not measured. misses in the shued case from 248% to 33%. more. Comparing AVL and red-black tree results, the AVL trees stricter balancing rule pays o, producing a tree of identiers with maximum path length of 14 (averaged over all insertions) instead of the red-black trees 25, and yielding the 13% to 16% edge of AVL trees in that category. 5.3.3 Shued Test Set

Discussion

The results for insertion in random order again reect the disadvantage of splay trees for random access and the advantage of the more relaxed red-black tree balancing rule. More curious is that every time in the shued test set is worse, by 13% to 21%, than the corresponding time in the normal set. Normally one would expect that the shued times should be about the same or better than the times for normal insertion order, because random order is ideal for BST insertion. The processor cache, along with the structure of the input data, turns out to be the culprit. When any given identier is used once in a source le, it tends to be used again within a few lines of code. This natural clustering yields a good hit rate in the processor cache. Shuing the input data destroys clustering and reduces the eect of the processor cache for large input data sets. To test this hypothesis, the experiment was rerun using only the rst 10,000 identiers from the test set. The results, shown in Table 7, conrm the idea: the edge of the normal input data over the shued input data is reduced to at most 7% for non-splay variants. Further investigation using the

The preceding sections examined three experiments with binary search tree-based data structures and discussed the results for each one in isolation. This section attempts to extend these individual results into broader conclusions about the performance of BSTbased data structures in a more general context. Each of the experiments involves three dierent kinds of test sets: those drawn from real-world situations, and variations that yield best- and worstcase performance in unbalanced BSTs. (For the peer cache experiment, the normal case is close to the best case.) The experimental results allow straightforward guidelines for eective choice of data structure to be drawn up, described in detail in the following sections.

6.1

Choice of Data Structure

In the best case for unbalanced BSTs, that is, insertion of items in random order, unbalanced trees are the best choice because of their low time and space overhead. If items are normally inserted in random order but include occasional runs in sorted order, then red-black trees are preferred to AVL trees because their more relaxed balancing rule does less work attempting to balance already random data. Splay trees are undesirable, because of the cost of splaying on every access. In the worst case for unbalanced BSTs, that is, insertion of items in sorted order, splay trees are the

10

best choice as long as subsequent accesses tend to be somewhat sequential (as in the VMA experiment in section 5.1) or clustered (as in the cross-reference experiment in section 5.3). Splay trees should not, however, be used if bounded performance guarantees are required. Otherwise, AVL trees are a better choice because they tend to keep the maximum path length below that for red-black trees, although the internal path length for AVL and red-black trees is comparable. Unbalanced BSTs should not be considered. Real-world situations fall somewhere between the two extremes: in two of our experiments red-black performance was better than AVL performance and in the other one it was the reverse; in one, splay trees were better than either. The choice of data structure can then be made based on which extreme is thought more likely or for which performance is more important, or based on actual tests.

case k = 1, but did not compare HB[k] trees to other forms of balanced tree. Baer and Schwab [2] empirically compared height-balanced, weight-balanced, and periodically totally restructured trees, using synthetic test sets, and concluded that AVL trees oer the best performance. Cannady [25] made theoretical comparisons height-balanced, bounded-balance, and weight-balanced trees. In 1980, Wright [3] analyzed the performance of several types of trees using a variety of synthetic test sets. One FreeBSD developer has reported that switching from doubly linked lists to splay trees increased the overall speed of a web server execution by 23% [26].

Conclusion

6.2

Choice of Node Representation

The results also aid the choice of BST representation. BST representations with parent pointers had the best performance in all three of our experiments, followed closely by threaded representations. Threads use less memory than a parent pointer and are therefore preferable when space is at a premium. A plain BST representation, without parent pointers or threads, saves time taken updating these additional elds, but the advantage is more than made up by costs incurred when traversal is combined with modications. A plain representation also seems to preclude ecient implementation of t delete. The linked list node representation was a clear winner for three out of four data sets for the VMA experiment, which made heavy use of traversal operations t next and t prev, but its use exacted a signicant penalty for the peer cache experiment. Its high memory cost, two pointer elds per nodes, means that it should only be used when traversal is common. Right threads seem rarely preferable, especially since there seems no ecient way to implement t delete. (Of course, t prev is also slow in rightthreaded trees.) Right threads performed poorly in all three experiments.

Related Work

Few comparisons of BST-based data structures exist in the literature, and none make use of real-world test sets or include red-black or splay trees. In 1976, Karlton et al. [1] experimentally examined the performance of HB[k] trees, a class of height-balanced binary trees for which AVL trees form the special

In this paper, we empirically compared the performance of 20 variants on BSTs in three dierent realworld scenarios, and demonstrated that choice of data structure can signicantly impact performance. Our results show that BST data structures and node representations should be chosen based on expected patterns in the input and the mix of operations to be performed. We found that in selecting data structures, unbalanced BSTs are best when randomly ordered input can be relied upon; if random ordering is the norm but occasional runs of sorted order are expected, then red-black trees should be chosen. On the other hand, if insertions often occur in a sorted order, AVL trees excel when later accesses tend to be random, and splay trees perform best when later accesses are sequential or clustered. For node representation, we found that parent pointers are generally fastest, so they should be preferred as long as the cost of an additional pointer eld per node is not important. If space is at a premium, threaded representations conserve memory and lag only slightly behind parent pointers in speed. A plain BST has fewer elds to update, but combining traversal and modication requires extra searches. Maintain a linked list of in-order nodes is a good choice when traversal is very common, but exacts a high memory cost. Finally, right-threaded representations fared poorly in all of our experiments. We also showed that implementing an routine for deleting a node referenced by a traverser can yield signicant performance improvement for some workloads. Based on our results future versions of libavl are likely to add support for such an operation, as well as splay trees.

11

Acknowledgments

[12] R. E. Tarjan, Sequential access in splay trees takes linear time, Combinatorica, vol. 5, no. 4, The author thanks advisor Mendel Rosenblum and pp. 367378, 1985. colleague Tal Garnkel. The author was supported by a Stanford Graduate Fellowship during the course [13] D. E. Knuth, Fundamental Algorithms, vol. 1 of The Art of Computer Programming, section of this work. 2.3.1, pp. 322327. Reading, Massachusetts: Addison-Wesley, third ed., 1997.

References

[14] B. Pfa, An Introduction to Binary Search Trees and Balanced Trees, vol. 1 of libavl Binary [1] P. L. Karlton, S. H. Fuller, R. E. Scroggs, and Search Tree Library. Free Software Foundation, E. B. Kaehler, Performance of height-balanced 2.0.1 ed., 2002. trees, Communications of the ACM, vol. 19, no. 1, pp. 2328, 1976. [15] M. K. McKusick, K. Bostic, M. J. Karels, and [2] J.-L. Baer and B. Schwab, A comparison of tree-balancing algorithms, Communications of the ACM, vol. 20, no. 5, pp. 322330, 1977. J. S. Quarterman, The design and implementation of the 4.4BSD operating system, ch. 5, pp. 117190. Addison Wesley Longman Publishing Co., Inc., 1996.

[3] W. E. Wright, An empirical evaluation of al[16] H. Custer, ed., Inside Windows NT, p. 200. Migorithms for dynamically maintaining binary crosoft Press, 1993. search trees, in Proceedings of the ACM 1980 annual conference, pp. 505515, 1980. [17] T. H. Cormen, C. E. Leiserson, and R. L. Rivest, Introduction to Algorithms, section 15.3, [4] D. E. Knuth, Sorting and Searching, vol. 3 of The pp. 290295. MIT Press, 1990. Art of Computer Programming, section 6.2.2, pp. 43031. Reading, Massachusetts: Addison- [18] Mozilla 1.0. Software, June 2002. Wesley, second ed., 1997. [19] VMware, Inc., VMware GSX Server 2.0.1. Software, 2002. [5] T. H. Cormen, C. E. Leiserson, and R. L. Rivest, Introduction to Algorithms, section 13, p. 244. [20] J. Postel, RFC 791: Internet Protocol, Sept. MIT Press, 1990. 1981. Status: STANDARD. [6] R. Sedgewick, Algorithms in C, Parts 14, sec[21] A. V. Savochkin, net/ipv4/inetpeer.c. Linux tion 12.6, pp. 508511. Addison-Wesley, 3rd ed., 2.4.18 kernel source, 2001. 1998. [22] T. H. Cormen, C. E. Leiserson, and R. L. [7] G. M. Adelson-Velski and E. M. Landis, An Rivest, Introduction to Algorithms, section algorithm for the organization of information, 12.3.3, pp. 229232. MIT Press, 1990. Soviet Mathematics Doklady, vol. 3, pp. 1259 1262, 1962. [23] International Organization for Standardization, ISO/IEC 14882:1998: Programming languages [8] D. E. Knuth, Sorting and Searching, vol. 3 C++. Geneva, Switzerland: International of The Art of Computer Programming, section Organization for Standardization, Sept. 1998. 6.2.3, p. 460. Reading, Massachusetts: Addison[24] Intel Corporation, Intel Architecture Software Wesley, second ed., 1997. Developer?s Manual Volume 3: System Pro[9] T. H. Cormen, C. E. Leiserson, and R. L. Rivest, gramming, 2002. Introduction to Algorithms, section 14.1, p. 264. [25] J. M. Cannady, Balancing methods for binary MIT Press, 1990. search trees, in Proceedings of the 16th an[10] R. Sedgewick, Algorithms in C, Parts 14, secnual conference on Southeast regional confertion 13.4, p. 556. Addison-Wesley, 3rd ed., 1998. ence, pp. 181186, ACM Press, 1978. [11] D. D. Sleator and R. E. Tarjan, Self-adjusting [26] A. Cox, fa.m7rv3cv.klm8jr@i.uio.no. Usenet, May 2002. binary search trees, Journal of the ACM, vol. 32, pp. 652686, July 1985. 12