The rapid growth of AI workloads at Meta has motivated our inhouse development of AI chips, aiming to significantly reduce the total cost of ownership and mitigate risks posed by unpredictable GPU supplies.At ISCA'23, we presented Meta's first-generation AI chip, MTIA 1.This paper describes its successor, MTIA 2i, now deployed at scale and serving billions of users.MTIA 2i significantly improves upon MTIA 1, reducing total cost of ownership by 44% compared to GPUs while delivering competitive performance per watt.A key differentiator is its memory hierarchy: instead of costly HBM, it uses large SRAM alongside LPDDR.Although there has been a proliferation of publications on AI chips, they often focus on architectural design and overlook three critical aspects:(1) co-designing and optimizing ML models to work effectively with the AI chip; (2) demonstrating sufficient flexibility to support a wide range of models; and (3) during the productionization process, addressing challenges unanticipated or decisions deferred at design time, such as dealing with memory errors, safe overclocking, reducing provisioned power, and implementing real-time firmware updates to mitigate silicon design defects.A key contribution of this paper is sharing our experience with these aspects, based on our journey of productionizing MTIA 2i at scale.

The rapid growth of server virtualization has ignited a wide adoption of software-based virtual switches, with significant interest in speeding up their performance. In a similar trend, software-defined networking (SDN), with its strong reliance on rule-based flow classification, has also created renewed interest in multi-dimensional packet classification. However, despite these recent advances, the performance of current software-based packet classifiers is still limited, mostly by the low parallelism of general-purpose CPUs. In this paper, we explore how to accelerate packet classification using the high parallelism and latency-hiding capabilities of graphic processing units (GPUs). We implement GPU-accelerated versions for both linear and tuple search, currently deployed in virtual switches, and also introduce a novel algorithm called Bloom search. These algorithms are integrated with high-speed packet I/O to build GSwitch, a GPU-accelerated software switch, and also to extend Open vSwitch. Our experimental evaluation indicates that, under realistic rule sets, GSwitch is at least 7 × faster than an equally-priced CPU classifier. We also show that our GPU-accelerated Open vSwitch outperforms the classic Open vSwitch implementation by a factor of 10, on average.

Network topology plays a critical role while designing and evaluating network protocols. Most existing topology generators are insufficient to reflect the real world network demands to a topology or to capture the Internet topology evolution such as the "flattening" Internet. They focused on the graph properties of a topology, thus, lacking of ability to model engineering features of the network. Some state-of-art topology generators that consider engineering factors fail to capture trends in both intra-AS and inter-AS connections, which are equally important for evaluating future network protocols. We have developed a topology generator GeoTopo, which is to our best knowledge the first scalable topology generator modeling engineering factors for both intra-AS and inter-AS topology generation. The engineering factors that GeoTopo considers include demographic and geographic features as well as business interests of ASes. We use GeoTopo to create and study three classes of topologies: the topology characterized mainly by graph-properties (Status Quo topology), the topology driven by peering at Internet Exchange Points (IXP topology) and the topology characterized by country backbones (CB topology). The SQ topology follows the degree-based model and serves as a baseline for capturing topology features. The IXP and CB topologies model two major directions of the Internet "flattening". The three classes of topologies enable us to analyze the impact of engineering factors on topology generation such as AS peering policies, IXP deployment and AS geo-settings. GeoTopo's ability to generate projected future Internet topologies make it a valuable tool for the design and evaluation of Future Internet Architectures that is currently under consideration in the research community. We use the evaluation of Global Name Resolution Service (GNRS), a key component shared by name-based network architectures, as an example application to demonstrate GeoTopo's capability to capture the mobility of network entities, the locality of the traffic, and the impact of the evolving network.

The information-centric networking (ICN) concept has been investigated to support mobile content delivery, in which content is distributed and requested directly through their names. Pervasive in-network caching is normally enabled in ICN to further facilitate mobile content delivery. Several different architectures have been proposed to realize ICN. In this work, we present their common features and then compare different approaches for mobile content delivery in these architectures. Such comparison would greatly deepen our understanding of different design choices and help with future ICN research.

The prevalence of mobile devices especially smartphones has attracted research on mobile content delivery techniques. In this paper, we propose to take advantage of the storage available at wireless access points to bring content closer to mobile devices, hence improving the downloading performance. Specifically, we propose to have a separate popularity based cache and a prefetch buffer at the network edge to capture both long-term and short-term content access patterns. Further, we point out that it is insufficient to rely on a device's past history to predict when and where to prefetch, especially in urban settings; instead, we propose to derive a prediction model based on the aggregated network-level statistics. We discuss the proposed mobile content caching/prefetching method in the context of the MobilityFirst future Internet architecture. In MobilityFirst, when mobile clients move between network attachment points (e.g., Wi-Fi access points), their network association records are logged by the network, which then naturally facilitates the network-level mobility prediction. Through detailed simulations with real taxi mobility traces, we show that such a strategy is more effective than earlier schemes in satisfying content requests at the edge (higher cache hit ratios), leading to shorter content download latencies. Specifically, the fraction of requests satisfied at the edge increases by a factor of 2.9 compared to a caching only approach, and by 45% compared to individual user-based prediction and prefetching.

The rapid growth of server virtualization has ignited a wide adoption of software-based virtual switches, with significant interest in speeding up their performance. In a similar trend, software-defined networking (SDN), with its strong reliance on rule-based flow classification, has also created renewed interest in multi-dimensional packet classification. However, despite these recent advances, the performance of current software-based packet classifiers is still limited, mostly by the low parallelism of general-purpose CPUs. In this paper, we explore how to accelerate packet classification using the high parallelism and latency-hiding capabilities of graphic processing units (GPUs). We implement GPU-accelerated versions for both linear and tuple search, currently deployed in virtual switches, and also introduce a novel algorithm called Bloom search. These algorithms are integrated with high-speed packet I/O to build GSwitch, a GPU-accelerated software switch. Our experimental evaluation shows that GSwitch is at least 7x faster than an equally-priced CPU classifier and is able to reach 10 Gbps with minimum-sized packets and a rule set containing 128K OpenFlow entries with 512 different wildcard patterns.

Content-centric networking (CCN) adopts a receiver-driven, hop-by-hop transport approach that facilitates in-network caching, which in turn leads to multiple sources and multiple paths for transferring content. In such a case, keeping a single round trip time (RTT) estimator for a multi-path flow is insufficient as each path may experience different round trip times. To solve this problem, it has been proposed to use multiple RTT estimators to predict network condition. In this paper, we examine an alternative approach to this problem, CHoPCoP, which utilizes explicit congestion control to cope with the multiple-source, multiple-path situation. Protocol design innovations of CHoPCoP include a random early marking (REM) scheme that explicitly signals network congestion, and a per-hop fair share Interest shaping algorithm (FISP) and a receiver Interest control method (RIC) that regulate the Interest rates at routers and the receiver respectively. We have implemented CHoPCoP on the ORBIT testbed and conducted experiments under various network and traffic settings. The evaluation shows that CHoPCoP is a viable approach that can effectively deal with congestion in the multipath environment.

The prevalence of mobile devices has attracted research on mobile content delivery techniques. The MobilityFirst(MF) project, discussed in this paper, proposes a clean-slate Internet architecture that enables mobile content-oriented operations at the network level. We describe the design details of the architecture in realizing this and provides a preliminary evaluation on scalability and performance.

Radio frequency based device-free passive (DfP) localization techniques have shown great potentials in localizing individual human subjects, without requiring them to carry any radio devices. In this study, we extend the DfP technique to count and localize multiple subjects in indoor environments. To address the impact of multipath on indoor radio signals, we adopt a fingerprinting based approach to infer subject locations from observed signal strengths through profiling the environment. When multiple subjects are present, our objective is to use the profiling data collected by a single subject to count and localize multiple subjects without any extra effort. In order to address the non-linearity of the impact of multiple subjects, we propose a successive cancellation based algorithm to iteratively determine the number of subjects. We model indoor human trajectories as a state transition process, exploit indoor human mobility constraints and integrate all information into a conditional random field (CRF) to simultaneously localize multiple subjects. As a result, we call the proposed algorithm SCPL -- sequential counting, parallel localizing. We test SCPL with two different indoor settings, one with size 150 m2 and the other 400 m2. In each setting, we have four different subjects, walking around in the deployed areas, sometimes with overlapping trajectories. Through extensive experimental results, we show that SCPL can count the present subjects with 86% counting percentage when their trajectories are not completely overlapping. Our localization algorithms are also highly accurate, with an average localization error distance of 1.3 m.

Multiple MIMO (Multiple Input Multiple Output) modes are the salient characteristics of 802.11n wireless networks. An efficient MIMO mode switching scheme is highly desirable for rate adaptation, which can optimize combination of MIMO mode and MCS (Modulation Coding Scheme) to achieve the maximum system throughput. In this paper, we have proposed and verified three design guidelines for the MIMO mode switching through real experiments. Based on these guidelines, we then present a simple and efficient MImo mode Switching Scheme, called MISS, and implement it into the Atheros-based device driver. MISS uses a crossover step-based searching algorithm to find the appropriate MIMO mode for rate adaptation. The efficiency of MISS has been evaluated through real experiments conducted in the indoor environments.