1 \documentclass[smallextended]{svjour3}
2 \usepackage[utf8]{inputenc}
3 \usepackage[T1]{fontenc}
9 \newcommand{\abs}[1]{\lvert#1\rvert} % \abs{x} -> |x|
13 \title{Best effort strategy and virtual load
14 for asynchronous iterative load balancing}
16 \author{Raphaël Couturier \and
21 \institute{R. Couturier \and A. Giersch \at
22 LIFC, University of Franche-Comté, Belfort, France \\
23 % Tel.: +123-45-678910\\
24 % Fax: +123-45-678910\\
26 raphael.couturier@univ-fcomte.fr,
27 arnaud.giersch@univ-fcomte.fr}
30 University of Béjaïa, Béjaïa, Algeria \\
31 \email{ar.sider@univ-bejaia.dz}
39 Most of the time, asynchronous load balancing algorithms have extensively been
40 studied in a theoretical point of view. The Bertsekas and Tsitsiklis'
41 algorithm~\cite[section~7.4]{bertsekas+tsitsiklis.1997.parallel}
42 is certainly the most well known algorithm for which the convergence proof is
43 given. From a practical point of view, when a node wants to balance a part of
44 its load to some of its neighbors, the strategy is not described. In this
45 paper, we propose a strategy called \emph{best effort} which tries to balance
46 the load of a node to all its less loaded neighbors while ensuring that all the
47 nodes concerned by the load balancing phase have the same amount of load.
48 Moreover, asynchronous iterative algorithms in which an asynchronous load
49 balancing algorithm is implemented most of the time can dissociate messages
50 concerning load transfers and message concerning load information. In order to
51 increase the converge of a load balancing algorithm, we propose a simple
52 heuristic called \emph{virtual load} which allows a node that receives an load
53 information message to integrate the load that it will receive later in its
54 load (virtually) and consequently sends a (real) part of its load to some of its
55 neighbors. In order to validate our approaches, we have defined a simulator
56 based on SimGrid which allowed us to conduct many experiments.
61 \section{Introduction}
63 Load balancing algorithms are extensively used in parallel and distributed
64 applications in order to reduce the execution times. They can be applied in
65 different scientific fields from high performance computation to micro sensor
66 networks. They are iterative by nature. In literature many kinds of load
67 balancing algorithms have been studied. They can be classified according
68 different criteria: centralized or decentralized, in static or dynamic
69 environment, with homogeneous or heterogeneous load, using synchronous or
70 asynchronous iterations, with a static topology or a dynamic one which evolves
71 during time. In this work, we focus on asynchronous load balancing algorithms
72 where computer nodes are considered homogeneous and with homogeneous load with
73 no external load. In this context, Bertsekas and Tsitsiklis have proposed an
74 algorithm which is definitively a reference for many works. In their work, they
75 proved that under classical hypotheses of asynchronous iterative algorithms and
76 a special constraint avoiding \emph{ping-pong} effect, an asynchronous
77 iterative algorithm converge to the uniform load distribution. This work has
78 been extended by many authors. For example, Cortés et al., with
79 DASUD~\cite{cortes+ripoll+cedo+al.2002.asynchronous}, propose a
80 version working with integer load. This work was later generalized by
81 the same authors in \cite{cedo+cortes+ripoll+al.2007.convergence}.
82 {\bf Rajouter des choses ici}.
84 Although the Bertsekas and Tsitsiklis' algorithm describes the condition to
85 ensure the convergence, there is no indication or strategy to really implement
86 the load distribution. In other word, a node can send a part of its load to one
87 or many of its neighbors while all the convergence conditions are
88 followed. Consequently, we propose a new strategy called \emph{best effort}
89 that tries to balance the load of a node to all its less loaded neighbors while
90 ensuring that all the nodes concerned by the load balancing phase have the same
91 amount of load. Moreover, when real asynchronous applications are considered,
92 using asynchronous load balancing algorithms can reduce the execution
93 times. Most of the times, it is simpler to distinguish load information messages
94 from data migration messages. Formers ones allows a node to inform its
95 neighbors of its current load. These messages are very small, they can be sent
96 quite often. For example, if an computing iteration takes a significant times
97 (ranging from seconds to minutes), it is possible to send a new load information
98 message at each neighbor at each iteration. Latter messages contains data that
99 migrates from one node to another one. Depending on the application, it may have
100 sense or not that nodes try to balance a part of their load at each computing
101 iteration. But the time to transfer a load message from a node to another one is
102 often much more longer that to time to transfer a load information message. So,
103 when a node receives the information that later it will receive a data message,
104 it can take this information into account and it can consider that its new load
105 is larger. Consequently, it can send a part of it real load to some of its
106 neighbors if required. We call this trick the \emph{virtual load} mechanism.
110 So, in this work, we propose a new strategy for improving the distribution of
111 the load and a simple but efficient trick that also improves the load
112 balancing. Moreover, we have conducted many simulations with SimGrid in order to
113 validate our improvements are really efficient. Our simulations consider that in
114 order to send a message, a latency delays the sending and according to the
115 network performance and the message size, the time of the reception of the
118 In the following of this paper, Section~\ref{BT algo} describes the Bertsekas
119 and Tsitsiklis' asynchronous load balancing algorithm. Moreover, we present a
120 possible problem in the convergence conditions. Section~\ref{Best-effort}
121 presents the best effort strategy which provides an efficient way to reduce the
122 execution times. In Section~\ref{Virtual load}, the virtual load mechanism is
123 proposed. Simulations allowed to show that both our approaches are valid using a
124 quite realistic model detailed in Section~\ref{Simulations}. Finally we give a
125 conclusion and some perspectives to this work.
130 \section{Bertsekas and Tsitsiklis' asynchronous load balancing algorithm}
133 In order prove the convergence of asynchronous iterative load balancing
134 Bertsekas and Tsitsiklis proposed a model
135 in~\cite{bertsekas+tsitsiklis.1997.parallel}. Here we recall some notations.
136 Consider that $N={1,...,n}$ processors are connected through a network.
137 Communication links are represented by a connected undirected graph $G=(N,V)$
138 where $V$ is the set of links connecting different processors. In this work, we
139 consider that processors are homogeneous for sake of simplicity. It is quite
140 easy to tackle the heterogeneous case~\cite{ElsMonPre02}. Load of processor $i$
141 at time $t$ is represented by $x_i(t)\geq 0$. Let $V(i)$ be the set of
142 neighbors of processor $i$. Each processor $i$ has an estimate of the load of
143 each of its neighbors $j \in V(i)$ represented by $x_j^i(t)$. According to
144 asynchronism and communication delays, this estimate may be outdated. We also
145 consider that the load is described by a continuous variable.
147 When a processor send a part of its load to one or some of its neighbors, the
148 transfer takes time to be completed. Let $s_{ij}(t)$ be the amount of load that
149 processor $i$ has transferred to processor $j$ at time $t$ and let $r_{ij}(t)$ be the
150 amount of load received by processor $j$ from processor $i$ at time $t$. Then
151 the amount of load of processor $i$ at time $t+1$ is given by:
153 x_i(t+1)=x_i(t)-\sum_{j\in V(i)} s_{ij}(t) + \sum_{j\in V(i)} r_{ji}(t)
158 Some conditions are required to ensure the convergence. One of them can be
159 called the \emph{ping-pong} condition which specifies that:
161 x_i(t)-\sum _{k\in V(i)} s_{ik}(t) \geq x_j^i(t)+s_{ij}(t)
163 for any processor $i$ and any $j \in V(i)$ such that $x_i(t)>x_j^i(t)$. This
164 condition aims at avoiding a processor to send a part of its load and being
165 less loaded after that.
167 Nevertheless, we think that this condition may lead to deadlocks in some
168 cases. For example, if we consider only three processors and that processor $1$
169 is linked to processor $2$ which is also linked to processor $3$ (i.e. a simple
170 chain which 3 processors). Now consider we have the following values at time $t$:
177 In this case, processor $2$ can either sends load to processor $1$ or processor
178 $3$. If it sends load to processor $1$ it will not satisfy condition
179 (\ref{eq:ping-pong}) because after the sending it will be less loaded that
180 $x_3^2(t)$. So we consider that the \emph{ping-pong} condition is probably to
181 strong. Currently, we did not try to make another convergence proof without this
182 condition or with a weaker condition.
185 \section{Best effort strategy}
188 We will describe here a new load-balancing strategy that we called
189 \emph{best effort}. The general idea behind this strategy is, for a
190 processor, to send some load to the most of its neighbors, doing its
191 best to reach the equilibrium between those neighbors and himself.
193 More precisely, when a processors $i$ is in its load-balancing phase,
194 he proceeds as following.
196 \item First, the neighbors are sorted in non-decreasing order of their
197 known loads $x^i_j(t)$.
199 \item Then, this sorted list is traversed in order to find its largest
200 prefix such as the load of each selected neighbor is lesser than:
202 \item the processor's own load, and
203 \item the mean of the loads of the selected neighbors and of the
206 Let's call $S_i(t)$ the set of the selected neighbors, and
207 $\bar{x}(t)$ the mean of the loads of the selected neighbors and of
210 \bar{x}(t) = \frac{1}{\abs{S_i(t)} + 1}
211 \left( x_i(t) + \sum_{j\in S_i(t)} x^i_j(t) \right)
213 The following properties hold:
216 S_i(t) \subset V(i) \\
217 x^i_j(t) < x_i(t) & \forall j \in S_i(t) \\
218 x^i_j(t) < \bar{x} & \forall j \in S_i(t) \\
219 x^i_j(t) \leq x^i_k(t) & \forall j \in S_i(t), \forall k \in V(i) \setminus S_i(t) \\
224 \item Once this selection is completed, processor $i$ sends to each of
225 the selected neighbor $j\in S_i(t)$ an amount of load $s_{ij}(t) =
228 From the above equations, and notably from the definition of
229 $\bar{x}$, it can easily be verified that:
232 x_i(t) - \sum_{j\in S_i(t)} s_{ij}(t) = \bar{x} \\
233 x^i_j(t) + s_{ij}(t) = \bar{x} & \forall j \in S_i(t)
238 \section{Other strategies}
241 \textbf{Question} faut-il décrire les stratégies makhoul et simple ?
243 \paragraph{simple} Tentative de respecter simplement les conditions de Bertsekas.
244 Parmi les voisins moins chargés que soi, on sélectionne :
246 \item un des moins chargés (vmin) ;
247 \item un des plus chargés (vmax),
249 puis on équilibre avec vmin en s'assurant que notre charge reste
250 toujours supérieure à celle de vmin et à celle de vmax.
252 On envoie donc (avec "self" pour soi-même) :
254 \min\left(\frac{load(self) - load(vmin)}{2}, load(self) - load(vmax)\right)
257 \paragraph{makhoul} Ordonne les voisins du moins chargé au plus chargé
258 puis calcule les différences de charge entre soi-même et chacun des
261 Ensuite, pour chaque voisin, dans l'ordre, et tant qu'on reste plus
262 chargé que le voisin en question, on lui envoie 1/(N+1) de la
263 différence calculée au départ, avec N le nombre de voisins.
265 C'est l'algorithme~2 dans~\cite{bahi+giersch+makhoul.2008.scalable}.
267 \section{Virtual load}
270 \section{Simulations}
273 In order to test and validate our approaches, we wrote a simulator
275 framework~\cite{casanova+legrand+quinson.2008.simgrid}. This
276 simulator, which consists of about 2,700 lines of C++, allows to run
277 the different load-balancing strategies under various parameters, such
278 as the initial distribution of load, the interconnection topology, the
279 characteristics of the running platform, etc. Then several metrics
280 are issued that permit to compare the strategies.
282 The simulation model is detailed in the next section (\ref{Sim
283 model}), then the results of the simulations are presented in
284 section~\ref{Results}.
286 \subsection{Simulation model}
289 In the simulation model the processors exchange messages which are of
290 two kinds. First, there are \emph{control messages} which only carry
291 information that is exchanged between the processors, such as the
292 current load, or the virtual load transfers if this option is
293 selected. These messages are rather small, and their size is
294 constant. Then, there are \emph{data messages} that carry the real
295 load transferred between the processors. The size of a data message
296 is a function of the amount of load that it carries, and it can be
297 pretty large. In order to receive the messages, each processor has
298 two receiving channels, one for each kind of messages. Finally, when
299 a message is sent or received, this is done by using the non-blocking
300 primitives of SimGrid\footnote{That are \texttt{MSG\_task\_isend()},
301 and \texttt{MSG\_task\_irecv()}.}.
303 During the simulation, each processor concurrently runs three threads:
304 a \emph{receiving thread}, a \emph{computing thread}, and a
305 \emph{load-balancing thread}, which we will briefly describe now.
307 \paragraph{Receiving thread} The receiving thread is in charge of
308 waiting for messages to come, either on the control channel, or on the
309 data channel. When a message is received, it is pushed in a buffer of
310 received message, to be later consumed by one of the other threads.
311 There are two such buffers, one for the control messages, and one for
312 the data messages. The buffers are implemented with a lock-free FIFO
313 \cite{sutter.2008.writing} to avoid contention between the threads.
315 \paragraph{Computing thread} The computing thread is in charge of the
316 real load management. It iteratively runs the following operations:
318 \item if some load was received from the neighbors, get it;
319 \item if there is some load to send to the neighbors, send it;
320 \item run some computation, whose duration is function of the current
321 load of the processor.
323 Practically, after the computation, the computing thread waits for a
324 small amount of time if the iterations are looping too fast (for
325 example, when the current load is zero).
327 \paragraph{Load-balancing thread} The load-balancing thread is in
328 charge of running the load-balancing algorithm, and exchange the
329 control messages. It iteratively runs the following operations:
331 \item get the control messages that were received from the neighbors;
332 \item run the load-balancing algorithm;
333 \item send control messages to the neighbors, to inform them of the
334 processor's current load, and possibly of virtual load transfers;
335 \item wait a minimum (configurable) amount of time, to avoid to
339 \subsection{Validation of our approaches}
343 On veut montrer quoi ? :
345 1) best plus rapide que les autres (simple, makhoul)
346 2) avantage virtual load
348 Est ce qu'on peut trouver des contre exemple?
352 Simulation avec temps définies assez long et on mesure la qualité avec : volume de calcul effectué, volume de données échangées
353 Mais aussi simulation avec temps court qui montre que seul best converge
356 Expés avec ratio calcul/comm rapide et lent
358 Quelques expés avec charge initiale aléatoire plutot que sur le premier proc
360 Cadre processeurs homogènes
364 On ne tient pas compte de la vitesse des liens donc on la considère homogène
366 Prendre un réseau hétérogène et rendre processeur homogène
368 Taille : 10 100 très gros
370 \section{Conclusion and perspectives}
373 \bibliographystyle{spmpsci}
374 \bibliography{biblio}
381 %%% ispell-local-dictionary: "american"
384 % LocalWords: Raphaël Couturier Arnaud Giersch Abderrahmane Sider Franche ij
385 % LocalWords: Bertsekas Tsitsiklis SimGrid DASUD Comté Béjaïa asynchronism ji
386 % LocalWords: ik isend irecv