Search

Low-Latency and Low-Power Edge AI is essential for Virtual Reality and Augmented Reality applications. Recent advances show that hybrid models, combining convolution layers (CNN) and transformers (ViT), often achieve superior accuracy/performance tradeoff on various computer vision and machine learning (ML) tasks. However, hybrid ML models can pose system challenges for latency and energy-efficiency due to their diverse nature in dataflow and memory access patterns. In this work, we leverage the architecture heterogeneity from Neural Processing Units (NPU) and Compute-In-Memory (CIM) and perform diverse execution schemas to efficiently execute these hybrid models. We also introduce H4H-NAS, a Neural Architecture Search framework to design efficient hybrid CNN/ViT models for heterogeneous edge systems with both NPU and CIM. Our H4H-NAS approach is powered by a performance estimator built with NPU performance results measured on real silicon, and CIM performance based on industry IPs. H4H-NAS searches hybrid CNN/ViT models with fine granularity and achieves significant (up to 1.34%) top-1 accuracy improvement on ImageNet dataset. Moreover, results from our Algo/HW co-design reveal up to 56.08% overall latency and 41.72% energy improvements by introducing such heterogeneous computing over baseline solutions. The framework guides the design of hybrid network architectures and system architectures of NPU+CIM heterogeneous systems.

Fine-tuning large language models (LLMs) requires significant memory, often exceeding the capacity of a single GPU. A common solution to this memory challenge is offloading compute and data from the GPU to the CPU. However, this approach is hampered by the limited bandwidth of commodity hardware, which constrains communication between the CPU and GPU. In this paper, we present an offloading framework, LSP_Offload, that enables near-native speed LLM fine-tuning on commodity hardware through learned subspace projectors. Our data-driven approach involves learning an efficient sparse compressor that minimizes communication with minimal precision loss. Additionally, we introduce a novel layer-wise communication schedule to maximize parallelism between communication and computation. As a result, our framework can fine-tune a 1.3 billion parameter model on a 4GB laptop GPU and a 7 billion parameter model on an NVIDIA RTX 4090 GPU with 24GB memory, achieving only a 31% slowdown compared to fine-tuning with unlimited memory. Compared to state-of-the-art offloading frameworks, our approach increases fine-tuning throughput by up to 3.33 times and reduces end-to-end fine-tuning time by 33.1%~62.5% when converging to the same accuracy.

We introduce RobotPerf, a vendor-agnostic benchmarking suite designed to evaluate robotics computing performance across a diverse range of hardware platforms using ROS 2 as its common baseline. The suite encompasses ROS 2 packages covering the full robotics pipeline and integrates two distinct benchmarking approaches: black-box testing, which measures performance by eliminating upper layers and replacing them with a test application, and grey-box testing, an application-specific measure that observes internal system states with minimal interference. Our benchmarking framework provides ready-to-use tools and is easily adaptable for the assessment of custom ROS 2 computational graphs. Drawing from the knowledge of leading robot architects and system architecture experts, RobotPerf establishes a standardized approach to robotics benchmarking. As an open-source initiative, RobotPerf remains committed to evolving with community input to advance the future of hardware-accelerated robotics.

Dynamic control flow is an important technique often used to design expressive and efficient deep learning computations for applications such as text parsing, machine translation, exiting early out of deep models and so on. The control flow divergence resulting from dynamic control flow makes batching, an important optimization enabling high throughput and hardware utilization, difficult to perform manually. In this paper, we present ACRoBat, a framework that enables efficient automatic batching for dynamic deep learning computations by performing hybrid static+dynamic compiler optimizations and end-to-end tensor code generation. ACRoBat performs up to 8.5X better than DyNet, a state-of-the-art framework for automatic batching, on an Nvidia GeForce GPU.

Batching has a fundamental influence on the efficiency of deep neural network (DNN) execution. However, for dynamic DNNs, efficient batching is particularly challenging as the dataflow graph varies per input instance. As a result, state-of-the-art frameworks use heuristics that result in suboptimal batching decisions. Further, batching puts strict restrictions on memory adjacency and can lead to high data movement costs. In this paper, we provide an approach for batching dynamic DNNs based on finite state machines, which enables the automatic discovery of batching policies specialized for each DNN via reinforcement learning. Moreover, we find that memory planning that is aware of the batching policy can save significant data movement overheads, which is automated by a PQ tree-based algorithm we introduce. Experimental results show that our framework speeds up state-of-the-art frameworks by on average 1.15x, 1.39x, and 2.45x for chain-based, tree-based, and lattice-based DNNs across CPU and GPU.

The performance of today's in-memory indexes is bottlenecked by the memory latency/bandwidth wall. Processing-in-memory (PIM) is an emerging approach that potentially mitigates this bottleneck, by enabling low-latency memory access whose aggregate memory bandwidth scales with the number of PIM nodes. There is an inherent tension, however, between minimizing inter-node communication and achieving load balance in PIM systems, in the presence of workload skew. This paper presents PIM-tree, an ordered index for PIM systems that achieves both low communication and high load balance, regardless of the degree of skew in the data and the queries. Our skew-resistant index is based on a novel division of labor between the multi-core host CPU and the PIM nodes, which leverages the strengths of each. We introduce push-pull search, which dynamically decides whether to push queries to a PIM-tree node (CPU -> PIM-node) or pull the node's keys back to the CPU (PIM-node -> CPU) based on workload skew. Combined with other PIM-friendly optimizations (shadow subtrees and chunked skip lists), our PIM-tree provides high-throughput, (guaranteed) low communication, and (guaranteed) high load balance, for batches of point queries, updates, and range scans. We implement the PIM-tree structure, in addition to prior proposed PIM indexes, on the latest PIM system from UPMEM, with 32 CPU cores and 2048 PIM nodes. On workloads with 500 million keys and batches of one million queries, the throughput using PIM-trees is up to 69.7x and 59.1x higher than the two best prior methods. As far as we know these are the first implementations of an ordered index on a real PIM system.

Federated Learning (FL) under distributed concept drift is a largely unexplored area. Although concept drift is itself a well-studied phenomenon, it poses particular challenges for FL, because drifts arise staggered in time and space (across clients). To the best of our knowledge, this work is the first to explicitly study data heterogeneity in both dimensions. We first demonstrate that prior solutions to drift adaptation that use a single global model are ill-suited to staggered drifts, necessitating multiple-model solutions. We identify the problem of drift adaptation as a time-varying clustering problem, and we propose two new clustering algorithms for reacting to drifts based on local drift detection and hierarchical clustering. Empirical evaluation shows that our solutions achieve significantly higher accuracy than existing baselines, and are comparable to an idealized algorithm with oracle knowledge of the ground-truth clustering of clients to concepts at each time step., Comment: 20 pages. Published in AISTATS 2023

Caches exploit temporal and spatial locality to allow a small memory to provide fast access to data stored in large, slow memory. The temporal aspect of locality is extremely well studied and understood, but the spatial aspect much less so. We seek to gain an increased understanding of spatial locality by defining and studying the Granularity-Change Caching Problem. This problem modifies the traditional caching setup by grouping data items into blocks, such that a cache can choose any subset of a block to load for the same cost as loading any individual item in the block. We show that modeling such spatial locality significantly changes the caching problem. This begins with a proof that Granularity-Change Caching is NP-Complete in the offline setting, even when all items have unit size and all blocks have unit load cost. In the online setting, we show a lower bound for competitive ratios of deterministic policies that is significantly worse than traditional caching. Moreover, we present a deterministic replacement policy called Item-Block Layered Partitioning and show that it obtains a competitive ratio close to that lower bound. Moreover, our bounds reveal a new issue arising in the Granularity-Change Caching Problem where the choice of offline cache size affects the competitiveness of different online algorithms relative to one another. To deal with this issue, we extend a prior (temporal) locality model to account for spatial locality, and provide a general lower bound in addition to an upper bound for Item-Block Layered Partitioning., Comment: 13 pages (including references), 6 figures, and 2 tables

There is often variation in the shape and size of input data used for deep learning. In many cases, such data can be represented using tensors with non-uniform shapes, or ragged tensors. Due to limited and non-portable support for efficient execution on ragged tensors, current deep learning frameworks generally use techniques such as padding and masking to make the data shapes uniform and then offload the computations to optimized kernels for dense tensor algebra. Such techniques can, however, lead to a lot of wasted computation and therefore, a loss in performance. This paper presents CoRa, a tensor compiler that allows users to easily generate efficient code for ragged tensor operators targeting a wide range of CPUs and GPUs. Evaluating CoRa on a variety of operators on ragged tensors as well as on an encoder layer of the transformer model, we find that CoRa (i)performs competitively with hand-optimized implementations of the operators and the transformer encoder and (ii) achieves, over PyTorch, a 1.6X geomean speedup for the encoder on an Nvidia GPU and a 1.86X geomean speedup for the multi-head attention module used in transformers on an ARM CPU., Comment: 23 pages, 25 figures and 10 tables

This paper introduces the first open-source FPGA-based infrastructure, MetaSys, with a prototype in a RISC-V core, to enable the rapid implementation and evaluation of a wide range of cross-layer techniques in real hardware. Hardware-software cooperative techniques are powerful approaches to improve the performance, quality of service, and security of general-purpose processors. They are however typically challenging to rapidly implement and evaluate in real hardware as they require full-stack changes to the hardware, OS, system software, and instruction-set architecture (ISA). MetaSys implements a rich hardware-software interface and lightweight metadata support that can be used as a common basis to rapidly implement and evaluate new cross-layer techniques. We demonstrate MetaSys's versatility and ease-of-use by implementing and evaluating three cross-layer techniques for: (i) prefetching for graph analytics; (ii) bounds checking in memory unsafe languages, and (iii) return address protection in stack frames; each technique only requiring ~100 lines of Chisel code over MetaSys. Using MetaSys, we perform the first detailed experimental study to quantify the performance overheads of using a single metadata management system to enable multiple cross-layer optimizations in CPUs. We identify the key sources of bottlenecks and system inefficiency of a general metadata management system. We design MetaSys to minimize these inefficiencies and provide increased versatility compared to previously-proposed metadata systems. Using three use cases and a detailed characterization, we demonstrate that a common metadata management system can be used to efficiently support diverse cross-layer techniques in CPUs., Comment: A shorter version of this work is to appear at the ACM Transactions on Architecture and Code Optimization (TACO). 27 pages, 15 figures

Optimizing deep learning models is generally performed in two steps: (i) high-level graph optimizations such as kernel fusion and (ii) low level kernel optimizations such as those found in vendor libraries. This approach often leaves significant performance on the table, especially for the case of recursive deep learning models. In this paper, we present Cortex, a compiler-based approach to generate highly-efficient code for recursive models for low latency inference. Our compiler approach and low reliance on vendor libraries enables us to perform end-to-end optimizations, leading to up to 14X lower inference latencies over past work, across different backends., Comment: 11 pages, 12 figures and 6 tables

When learning from streaming data, a change in the data distribution, also known as concept drift, can render a previously-learned model inaccurate and require training a new model. We present an adaptive learning algorithm that extends previous drift-detection-based methods by incorporating drift detection into a broader stable-state/reactive-state process. The advantage of our approach is that we can use aggressive drift detection in the stable state to achieve a high detection rate, but mitigate the false positive rate of standalone drift detection via a reactive state that reacts quickly to true drifts while eliminating most false positives. The algorithm is generic in its base learner and can be applied across a variety of supervised learning problems. Our theoretical analysis shows that the risk of the algorithm is competitive to an algorithm with oracle knowledge of when (abrupt) drifts occur. Experiments on synthetic and real datasets with concept drifts confirm our theoretical analysis., Comment: 32 pages, 12 figures. Submitted to NeurIPS 2020. Replaced to include revision of Lemma 2 and additional experimental results

Federated learning (FL) is a machine learning setting where many clients (e.g. mobile devices or whole organizations) collaboratively train a model under the orchestration of a central server (e.g. service provider), while keeping the training data decentralized. FL embodies the principles of focused data collection and minimization, and can mitigate many of the systemic privacy risks and costs resulting from traditional, centralized machine learning and data science approaches. Motivated by the explosive growth in FL research, this paper discusses recent advances and presents an extensive collection of open problems and challenges., Comment: Published in Foundations and Trends in Machine Learning Vol 4 Issue 1. See: https://www.nowpublishers.com/article/Details/MAL-083

Non-volatile main memory (NVRAM) technologies provide an attractive set of features for large-scale graph analytics, including byte-addressability, low idle power, and improved memory-density. NVRAM systems today have an order of magnitude more NVRAM than traditional memory (DRAM). NVRAM systems could therefore potentially allow very large graph problems to be solved on a single machine, at a modest cost. However, a significant challenge in achieving high performance is in accounting for the fact that NVRAM writes can be much more expensive than NVRAM reads. In this paper, we propose an approach to parallel graph analytics using the Parallel Semi-Asymmetric Model (PSAM), in which the graph is stored as a read-only data structure (in NVRAM), and the amount of mutable memory is kept proportional to the number of vertices. Similar to the popular semi-external and semi-streaming models for graph analytics, the PSAM approach assumes that the vertices of the graph fit in a fast read-write memory (DRAM), but the edges do not. In NVRAM systems, our approach eliminates writes to the NVRAM, among other benefits. To experimentally study this new setting, we develop Sage, a parallel semi-asymmetric graph engine with which we implement provably-efficient (and often work-optimal) PSAM algorithms for over a dozen fundamental graph problems. We experimentally study Sage using a 48-core machine on the largest publicly-available real-world graph (the Hyperlink Web graph with over 3.5 billion vertices and 128 billion edges) equipped with Optane DC Persistent Memory, and show that Sage outperforms the fastest prior systems designed for NVRAM. Importantly, we also show that Sage nearly matches the fastest prior systems running solely in DRAM, by effectively hiding the costs of repeatedly accessing NVRAM versus DRAM., Comment: This is an extended version of a paper in PVLDB (to be presented at VLDB'20)

Many large-scale machine learning (ML) applications need to perform decentralized learning over datasets generated at different devices and locations. Such datasets pose a significant challenge to decentralized learning because their different contexts result in significant data distribution skew across devices/locations. In this paper, we take a step toward better understanding this challenge by presenting a detailed experimental study of decentralized DNN training on a common type of data skew: skewed distribution of data labels across devices/locations. Our study shows that: (i) skewed data labels are a fundamental and pervasive problem for decentralized learning, causing significant accuracy loss across many ML applications, DNN models, training datasets, and decentralized learning algorithms; (ii) the problem is particularly challenging for DNN models with batch normalization; and (iii) the degree of data skew is a key determinant of the difficulty of the problem. Based on these findings, we present SkewScout, a system-level approach that adapts the communication frequency of decentralized learning algorithms to the (skew-induced) accuracy loss between data partitions. We also show that group normalization can recover much of the accuracy loss of batch normalization.

Machine learning (ML) techniques are enjoying rapidly increasing adoption. However, designing and implementing the systems that support ML models in real-world deployments remains a significant obstacle, in large part due to the radically different development and deployment profile of modern ML methods, and the range of practical concerns that come with broader adoption. We propose to foster a new systems machine learning research community at the intersection of the traditional systems and ML communities, focused on topics such as hardware systems for ML, software systems for ML, and ML optimized for metrics beyond predictive accuracy. To do this, we describe a new conference, MLSys, that explicitly targets research at the intersection of systems and machine learning with a program committee split evenly between experts in systems and ML, and an explicit focus on topics at the intersection of the two.

We consider a parallel computational model that consists of $P$ processors, each with a fast local ephemeral memory of limited size, and sharing a large persistent memory. The model allows for each processor to fault with bounded probability, and possibly restart. On faulting all processor state and local ephemeral memory are lost, but the persistent memory remains. This model is motivated by upcoming non-volatile memories that are as fast as existing random access memory, are accessible at the granularity of cache lines, and have the capability of surviving power outages. It is further motivated by the observation that in large parallel systems, failure of processors and their caches is not unusual. Within the model we develop a framework for developing locality efficient parallel algorithms that are resilient to failures. There are several challenges, including the need to recover from failures, the desire to do this in an asynchronous setting (i.e., not blocking other processors when one fails), and the need for synchronization primitives that are robust to failures. We describe approaches to solve these challenges based on breaking computations into what we call capsules, which have certain properties, and developing a work-stealing scheduler that functions properly within the context of failures. The scheduler guarantees a time bound of $O(W/P_A + D(P/P_A) \lceil\log_{1/f} W\rceil)$ in expectation, where $W$ and $D$ are the work and depth of the computation (in the absence of failures), $P_A$ is the average number of processors available during the computation, and $f \le 1/2$ is the probability that a capsule fails. Within the model and using the proposed methods, we develop efficient algorithms for parallel sorting and other primitives., Comment: This paper is the full version of a paper at SPAA 2018 with the same name

In existing systems, to perform any bulk data movement operation (copy or initialization), the data has to first be read into the on-chip processor, all the way into the L1 cache, and the result of the operation must be written back to main memory. This is despite the fact that these operations do not involve any actual computation. RowClone exploits the organization and operation of commodity DRAM to perform these operations completely inside DRAM using two mechanisms. The first mechanism, Fast Parallel Mode, copies data between two rows inside the same DRAM subarray by issuing back-to-back activate commands to the source and the destination row. The second mechanism, Pipelined Serial Mode, transfers cache lines between two banks using the shared internal bus. RowClone significantly reduces the raw latency and energy consumption of bulk data copy and initialization. This reduction directly translates to improvement in performance and energy efficiency of systems running copy or initialization-intensive workloads, Comment: arXiv admin note: text overlap with arXiv:1605.06483

The application resource specification--a static specification of several parameters such as the number of threads and the scratchpad memory usage per thread block--forms a critical component of modern GPU programming models. This specification determines the parallelism, and hence performance, of the application during execution because the corresponding on-chip hardware resources are allocated and managed based on this specification. This tight-coupling between the software-provided resource specification and resource management in hardware leads to significant challenges in programming ease, portability, and performance. Zorua is a new resource virtualization framework, that decouples the programmer-specified resource usage of a GPU application from the actual allocation in the on-chip hardware resources. Zorua enables this decoupling by virtualizing each resource transparently to the programmer. We demonstrate that by providing the illusion of more resources than physically available via controlled and coordinated virtualization, Zorua offers several important benefits: (i) Programming Ease. Zorua eases the burden on the programmer to provide code that is tuned to efficiently utilize the physically available on-chip resources. (ii) Portability. Zorua alleviates the necessity of re-tuning an application's resource usage when porting the application across GPU generations. (iii) Performance. By dynamically allocating resources and carefully oversubscribing them when necessary, Zorua improves or retains the performance of applications that are already highly tuned to best utilize the resources., Comment: arXiv admin note: substantial text overlap with arXiv:1802.02573

MLtuner automatically tunes settings for training tunables (such as the learning rate, the momentum, the mini-batch size, and the data staleness bound) that have a significant impact on large-scale machine learning (ML) performance. Traditionally, these tunables are set manually, which is unsurprisingly error-prone and difficult to do without extensive domain knowledge. MLtuner uses efficient snapshotting, branching, and optimization-guided online trial-and-error to find good initial settings as well as to re-tune settings during execution. Experiments show that MLtuner can robustly find and re-tune tunable settings for a variety of ML applications, including image classification (for 3 models and 2 datasets), video classification, and matrix factorization. Compared to state-of-the-art ML auto-tuning approaches, MLtuner is more robust for large problems and over an order of magnitude faster.

The application resource specification--a static specification of several parameters such as the number of threads and the scratchpad memory usage per thread block--forms a critical component of the existing GPU programming models. This specification determines the performance of the application during execution because the corresponding on-chip hardware resources are allocated and managed purely based on this specification. This tight coupling between the software-provided resource specification and resource management in hardware leads to significant challenges in programming ease, portability, and performance, as we demonstrate in this work. Our goal in this work is to reduce the dependence of performance on the software-provided resource specification to simultaneously alleviate the above challenges. To this end, we introduce Zorua, a new resource virtualization framework, that decouples the programmer-specified resource usage of a GPU application from the actual allocation in the on-chip hardware resources. Zorua enables this decoupling by virtualizing each resource transparently to the programmer. We demonstrate that by providing the illusion of more resources than physically available, Zorua offers several important benefits: (i) Programming Ease: Zorua eases the burden on the programmer to provide code that is tuned to efficiently utilize the physically available on-chip resources. (ii) Portability: Zorua alleviates the necessity of re-tuning an application's resource usage when porting the application across GPU generations. (iii) Performance: By dynamically allocating resources and carefully oversubscribing them when necessary, Zorua improves or retains the performance of applications that are already highly tuned to best utilize the resources. The holistic virtualization provided by Zorua has many other potential uses which we describe in this paper.

Large volumes of videos are continuously recorded from cameras deployed for traffic control and surveillance with the goal of answering "after the fact" queries: identify video frames with objects of certain classes (cars, bags) from many days of recorded video. While advancements in convolutional neural networks (CNNs) have enabled answering such queries with high accuracy, they are too expensive and slow. We build Focus, a system for low-latency and low-cost querying on large video datasets. Focus uses cheap ingestion techniques to index the videos by the objects occurring in them. At ingest-time, it uses compression and video-specific specialization of CNNs. Focus handles the lower accuracy of the cheap CNNs by judiciously leveraging expensive CNNs at query-time. To reduce query time latency, we cluster similar objects and hence avoid redundant processing. Using experiments on video streams from traffic, surveillance and news channels, we see that Focus uses 58X fewer GPU cycles than running expensive ingest processors and is 37X faster than processing all the video at query time.

The future of main memory appears to lie in the direction of new technologies that provide strong capacity-to-performance ratios, but have write operations that are much more expensive than reads in terms of latency, bandwidth, and energy. Motivated by this trend, we propose sequential and parallel algorithms to solve graph connectivity problems using significantly fewer writes than conventional algorithms. Our primary algorithmic tool is the construction of an $o(n)$-sized "implicit decomposition" of a bounded-degree graph $G$ on $n$ nodes, which combined with read-only access to $G$ enables fast answers to connectivity and biconnectivity queries on $G$. The construction breaks the linear-write "barrier", resulting in costs that are asymptotically lower than conventional algorithms while adding only a modest cost to querying time. For general non-sparse graphs on $m$ edges, we also provide the first $o(m)$ writes and $O(m)$ operations parallel algorithms for connectivity and biconnectivity. These algorithms provide insight into how applications can efficiently process computations on large graphs in systems with read-write asymmetry.

Bitwise operations are an important component of modern day programming. Many widely-used data structures (e.g., bitmap indices in databases) rely on fast bitwise operations on large bit vectors to achieve high performance. Unfortunately, in existing systems, regardless of the underlying architecture (e.g., CPU, GPU, FPGA), the throughput of such bulk bitwise operations is limited by the available memory bandwidth. We propose Buddy, a new mechanism that exploits the analog operation of DRAM to perform bulk bitwise operations completely inside the DRAM chip. Buddy consists of two components. First, simultaneous activation of three DRAM rows that are connected to the same set of sense amplifiers enables us to perform bitwise AND and OR operations. Second, the inverters present in each sense amplifier enables us to perform bitwise NOT operations, with modest changes to the DRAM array. These two components make Buddy functionally complete. Our implementation of Buddy largely exploits the existing DRAM structure and interface, and incurs low overhead (1% of DRAM chip area). Our evaluations based on SPICE simulations show that, across seven commonly-used bitwise operations, Buddy provides between 10.9X---25.6X improvement in raw throughput and 25.1X---59.5X reduction in energy consumption. We evaluate three real-world data-intensive applications that exploit bitwise operations: 1) bitmap indices, 2) BitWeaving, and 3) bitvector-based implementation of sets. Our evaluations show that Buddy significantly outperforms the state-of-the-art., Comment: arXiv admin note: text overlap with arXiv:1605.06483

Emerging memory technologies have a significant gap between the cost, both in time and in energy, of writing to memory versus reading from memory. In this paper we present models and algorithms that account for this difference, with a focus on write-efficient sorting algorithms. First, we consider the PRAM model with asymmetric write cost, and show that sorting can be performed in $O\left(n\right)$ writes, $O\left(n \log n\right)$ reads, and logarithmic depth (parallel time). Next, we consider a variant of the External Memory (EM) model that charges $\omega > 1$ for writing a block of size $B$ to the secondary memory, and present variants of three EM sorting algorithms (multi-way mergesort, sample sort, and heapsort using buffer trees) that asymptotically reduce the number of writes over the original algorithms, and perform roughly $\omega$ block reads for every block write. Finally, we define a variant of the Ideal-Cache model with asymmetric write costs, and present write-efficient, cache-oblivious parallel algorithms for sorting, FFTs, and matrix multiplication. Adapting prior bounds for work-stealing and parallel-depth-first schedulers to the asymmetric setting, these yield parallel cache complexity bounds for machines with private caches or with a shared cache, respectively.

In several emerging technologies for computer memory (main memory), the cost of reading is significantly cheaper than the cost of writing. Such asymmetry in memory costs poses a fundamentally different model from the RAM for algorithm design. In this paper we study lower and upper bounds for various problems under such asymmetric read and write costs. We consider both the case in which all but $O(1)$ memory has asymmetric cost, and the case of a small cache of symmetric memory. We model both cases using the $(M,\omega)$-ARAM, in which there is a small (symmetric) memory of size $M$ and a large unbounded (asymmetric) memory, both random access, and where reading from the large memory has unit cost, but writing has cost $\omega\gg 1$. For FFT and sorting networks we show a lower bound cost of $\Omega(\omega n\log_{\omega M} n)$, which indicates that it is not possible to achieve asymptotic improvements with cheaper reads when $\omega$ is bounded by a polynomial in $M$. Also, there is an asymptotic gap (of $\min(\omega,\log n)/\log(\omega M)$) between the cost of sorting networks and comparison sorting in the model. This contrasts with the RAM, and most other models. We also show a lower bound for computations on an $n\times n$ diamond DAG of $\Omega(\omega n^2/M)$ cost, which indicates no asymptotic improvement is achievable with fast reads. However, we show that for the edit distance problem (and related problems), which would seem to be a diamond DAG, there exists an algorithm with only $O(\omega n^2/(M\min(\omega^{1/3},M^{1/2})))$ cost. To achieve this we make use of a "path sketch" technique that is forbidden in a strict DAG computation. Finally, we show several interesting upper bounds for shortest path problems, minimum spanning trees, and other problems. A common theme in many of the upper bounds is to have redundant computation to tradeoff between reads and writes.