Modern data center networks support a wide range of applications with dynamic and diverse traffic patterns, necessitating effective load balancing to ensure performance and scalability. This paper presents ARION, a novel distributed load balancing algorithm designed for leaf-spine architectures. ARION utilizes flow aggregation and real-time link utilization metrics, enabling each leaf switch to make informed, congestionaware routing decisions without requiring any hardware modifications to the spine switches. This approach enables effective load distribution while substantially reducing packet reordering and associated overhead. We establish a theoretical framework to assess ARION’s performance, including convergence properties and a performance bound, demonstrating that it operates within a factor of 2 of the optimal solution under heavy loads. Comprehensive simulations covering various workloads and traffic types reveal that ARION consistently outperforms common alternatives such as ECMP, CONGA, DRILL, and LetFlow, achieving a superior balance between optimal load distribution and minimal packet reordering. ARION offers a scalable, low-overhead solution for the evolving demands of data center environments.

Transformer models have achieved remarkable results in a wide range of applications. However, their scalability is hampered by the quadratic time and memory complexity of the self-attention mechanism concerning the sequence length. This limitation poses a substantial obstacle when dealing with long documents or high-resolution images. In this work, we study the self-attention mechanism by analyzing the distribution of the attention matrix and its concentration ability. Furthermore, we propose instruments to measure these quantities and introduce a novel self-attention mechanism, Linear Log-Normal Attention, designed to emulate the distribution and concentration behavior of the original self-attention. Our experimental results on popular natural language benchmarks reveal that our proposed Linear Log-Normal Attention outperforms other linearized attention alternatives, offering a promising avenue for enhancing the scalability of transformer models.

Distributed scheduling is an attractive approach for the Multiple-Access Channel (MAC). However, when a subset of the users access the channel simultaneously, distributed rate coordination is necessary, and is a major challenge, since the achievable rate of each user highly depends on the channels of other active users. That is, given a detection technique, e.g., Zero-Forcing (ZF), the rate at which a user can transmit depends on the channels other transmitting users have, a knowledge which is usually unavailable in distributed schemes. Fixing a rate and accepting some outage probability when this rate is too high is common practice in these cases. In this paper, we analyze the outage probability of a distributed, asymptotically optimal threshold-based scheduling algorithm under ZF. We rigorously evaluate the distribution of the relevant projections and give upper and lower bounds on the outage probability as a function of the algorithm parameters. At the limit of a large number of users, the bounds match, resulting in the true asymptotic characterization of the outage probability.

The 802.11ac is a significant landmark in wireless communications, as it pushes towards new rate limits by utilizing downlink multiuser multiple-input multiple-output (MIMO) beamforming to transmit data to various locations simultaneously. However, successful beamforming relies on intelligent user selection which requires, in turn, extensive overhead of channel calibration between the AP and each of the candidate users. The large overhead involved in the user selection procedure overwhelms the multiuser gain and hinders the utilization of multiuser MIMO. The phenomenon is even more acute when APs handle large groups of mobile users, which frequently associate and disconnect, making the process of acquiring channel state from all users and selecting the appropriate group even harder. Thus, the subtle relation between the achievable rate of a scheduling algorithm and the overhead it requires is significant for the 802.11ac performance analysis. In this paper, we provide a rigorous analysis of distributed algorithms that schedule a group of users for the downlink. In particular, we accommodate common scheduling methods for the 802.11ac protocol and analyze both their achievable rate and their calibration process overhead. Both analysis and extensive simulations depict the superiority of simple threshold-based methods in terms of the throughput.

Distributed scheduling is an attractive approach for the Multiple-Access Channel (MAC). However, when a subset of the users access the channel simultaneously, distributed rate coordination is necessary, and is a major challenge, since the channel capacity of each user highly depends on the channels of other active users. That is, given a detection technique, e.g., Zero-Forcing (ZF), the rate at which a user can transmit depends on the channels other transmitting users have, a knowledge which is usually unavailable in distributed schemes. Fixing a rate and accepting some outage probability when this rate is too high is common practice in these cases. In this paper, we analyze the outage probability of a distributed, asymptotically optimal threshold-based scheduling algorithm under ZF. We rigorously evaluate the distribution of the relevant projections, and give upper and lower bounds on the outage probability as a function of the algorithm parameters.

Consider the problem of a multiple-input multiple-output multiple-access channel at the limit of large number of users. Clearly, in practical scenarios, only a small subset of the users can be scheduled to utilize the channel simultaneously. Thus, a problem of user selection arises. However, since solutions which collect channel state information from all users and decide on the best subset to transmit in each slot do not scale when the number of users is large, distributed algorithms for user selection are advantageous. In this paper, we analyze a distributed user selection algorithm, which selects a group of users to transmit without coordinating between users and without all users sending CSI to the base station. This threshold-based algorithm is analyzed for both zero-forcing and minimum mean square error receivers, and its expected sum rate in the limit of large number of users is investigated. It is shown that for large number of users, it achieves the same scaling laws as the optimal centralized scheme.

Network traffic classification is fundamental to network management and its performance. However, traditional approaches for traffic classification, which were designed to work on a dedicated hardware at very high line rates, may not function well in a virtual software-based environment. In this paper, we devise a novel fingerprinting technique that can be utilized as a software-based solution which enables machine-learning-based classification of ongoing flows. The suggested scheme is very simple to implement and requires minimal resources, yet attains very high accuracy. Specifically, for TCP flows, we suggest a fingerprint that is based on zero-length packets, hence enables a highly efficient sampling strategy which can be adopted with a single content-addressable memory rule. The suggested fingerprinting scheme is robust to network conditions such as congestion, fragmentation, delay, retransmissions, duplications, and losses and to varying processing capabilities. Hence, its performance is essentially independent of placement and migration issues, and thus yields an attractive solution for virtualized software-based environments. We suggest an analogous fingerprinting scheme for user datagram protocol traffic, which benefits from the same advantages as the TCP one and attains very high accuracy as well. Results show that our scheme correctly classified about 97% of the flows on the dataset tested, even on encrypted data.

In this paper, we study the secrecy rate and outage probability in Multiple-Input-Single-Output (MISO) Gaussian wiretap channels at the limit of a large number of legitimate users and eavesdroppers. In particular, we analyze the asymptotic achievable secrecy rates and outage, when only statistical knowledge on the wiretap channels is available to the transmitter. The analysis provides exact expressions for the reduction in the secrecy rate as the number of eavesdroppers grows, compared to the boost in the secrecy rate as the number of legitimate users grows.

In this work, we consider the problem of a Multiple-Input Single-Output (MISO) Multiple-Access Channel with a large number of users K. In practical scenarios, only a small sub-set of the users can be scheduled simultaneously. However, since solutions which collect Channel State Information (CSI) from all users and schedule the best subset to transmit do not scale with K, distributed scheduling algorithms are advantageous. We analyze a distributed scheduling algorithm, which selects a group of users to transmit without coordinating between the users and without all users sending CSI to the base station. The expected capacity under Minimum Mean Squared Error (MMSE) decoding is given, with a special emphasis on large K. It is shown that the algorithm achieves the same scaling laws as the optimal centralized scheme.

In this paper, we design novel distributed scheduling algorithms for multiuser multiple-input multiple-output systems and evaluate the resulting system capacity analytically. In particular, we consider algorithms which do not require sending channel state information to a central processing unit, nor do they require communication between the users themselves, yet, the resulting capacity closely approximates that of a centrally controlled system, which is able to schedule the strongest user in each time-slot. In other words, multiuser diversity is achieved in a distributed fashion. Our analysis is based on a novel application of the point-process approximation. This technique, besides tackling previously suggested models successfully, allows an analytical examination of new models, such as nonhomogeneous cases (nonidentically distributed users) or various quality of service considerations. This results in asymptotically exact expressions for the capacity of the system under these schemes, solving analytically problems which to date had been open. Possible applications include, but are not limited to, modern 4G networks, such as 3GPP LTE, or random access protocols.

Consider the problem of a multiple-antenna Multiple-Access Channel at the limit of large number of users. Clearly, in practical scenarios, only a small subset of the users can be scheduled to utilize the channel simultaneously. Thus, a problem of user selection arises. Since solutions which collect Channel State Information (CSI) from all users and decide on the best subset to transmit in each slot do not scale when the number of users is large, distributed algorithms for user selection are advantageous. In this paper, we suggest distributed user selection algorithms which select a group of users to transmit without coordinating between all users and without all users sending CSI to the base station. These threshold-based algorithms are analyzed, and their expected capacity in the limit of large number of users is investigated. It is shown that for large number of users a distributed algorithm can achieve the same scaling laws as the optimal centralized scheme.