Numerous applications today rely on artificial intelligence over images. Image AI is, however, extremely expensive. In particular, the inference cost of image AI dominates the end-to-end cost. We observe that the image storage format lies at the root of the problem. Images today are predominantly stored in JPEG format. JPEG is a storage format designed for the human eye; it maximally compresses images without distorting the components of an image that are visible to the human eye. However, our observation is that during image AI, images are "seen'' by algorithms, not humans. In addition, every AI application is different regarding which data components of the images are the most relevant. We present the Image Calculator, a self-designing image storage format that adapts to the given AI task, i.e., the specific neural network, the dataset, and the applications' specific accuracy, inference time, and storage requirements. Contrary to the state-of-the-art, the Image Calculator does not use a fixed storage format like JPEG. Instead, it designs and constructs a new storage format tailored to the context. It does so by constructing a massive design space of candidate storage formats from first principles, within which it searches efficiently using composite performance models (inference time, accuracy, storage). This way, it leverages the given AI task's unique characteristics to compress the data maximally. We evaluate the Image Calculator across a diverse set of data, image analysis tasks, AI models, and hardware. We show that the Image Calculator can generate image storage formats that reduce inference time by up to 14.2x and storage by up to 8.2x with a minimal loss in accuracy or gain, compared to JPEG and its state-of-the-art variants.

Hashing is a widely used technique for creating uniformly random numbers from arbitrary data. This is required in a large range of core data-driven operations including indexing, partitioning, filters, and sketches. As such, hashing is a core component in numerous systems including relational data systems, key-value stores, compilers, and networks. Due to both the computational and data heavy nature of hashing, it is a core systems bottleneck. For example, a typical database query in the standard TPC-H benchmark may spend 50% of its total cost in hash tables. Similarly, Google spends at least 2% of its total computational cost on C++ hash tables, resulting in a massive yearly cost footprint just from one hashing operation.

Hybrid Transactional and Analytical Processing (HTAP) systems have become popular in the past decade. HTAP systems allow running transactional and analytical processing workloads on the same data and hardware. As a result, they suffer from workload interference. Despite the large body of existing work in HTAP systems and architectures, none of the existing work has systematically analyzed workload interference for HTAP systems.In this work, we characterize workload interference for HTAP systems. We show that the OLTP throughput drops by up to 42% due to sharing the hardware resources. Partitioning the last-level cache (LLC) among the OLTP and OLAP workloads can significantly improve the OLTP throughput without hurting the OLAP throughput. The OLAP throughput is significantly reduced due to sharing the data. The OLAP execution time is exponentially increased if the OLTP workload generates fresh tuples faster than the HTAP system propagates them. Therefore, in order to minimize the workload interference, HTAP systems should isolate the OLTP and OLAP workloads in the shared hardware resources and should allocate enough resources to fresh tuple propagation to propagate the fresh tuples faster than they are generated.

Abstract Micro-architectural behavior of traditional disk-based online transaction processing (OLTP) systems has been investigated extensively over the past couple of decades. Results show that traditional OLTP systems mostly under-utilize the available micro-architectural resources. In-memory OLTP systems, on the other hand, process all the data in main-memory and, therefore, can omit the buffer pool. Furthermore, they usually adopt more lightweight concurrency control mechanisms, cache-conscious data structures, and cleaner codebases since they are usually designed from scratch. Hence, we expect significant differences in micro-architectural behavior when running OLTP on platforms optimized for in-memory processing as opposed to disk-based database systems. In particular, we expect that in-memory systems exploit micro-architectural features such as instruction and data caches significantly better than disk-based systems. This paper sheds light on the micro-architectural behavior of in-memory database systems by analyzing and contrasting it to the behavior of disk-based systems when running OLTP workloads. The results show that, despite all the design changes, in-memory OLTP exhibits very similar micro-architectural behavior to disk-based OLTP: more than half of the execution time goes to memory stalls where instruction cache misses or the long-latency data misses from the last-level cache (LLC) are the dominant factors in the overall execution time. Even though ground-up designed in-memory systems can eliminate the instruction cache misses, the reduction in instruction stalls amplifies the impact of LLC data misses. As a result, only 30% of the CPU cycles are used to retire instructions, and 70% of the CPU cycles are wasted to stalls for both traditional disk-based and new generation in-memory OLTP.

Understanding micro-architectural behavior is important for efficiently using hardware resources. Recent work has shown that in-memory online transaction processing (OLTP) systems severely underutilize their core micro-architecture resources [29]. Whereas, online analytical processing (OLAP) workloads exhibit a completely different computing pattern. OLAP workloads are read-only, bandwidth-intensive, and include various data access patterns. With the rise of column-stores, they run on high-performance engines that are tightly optimized for modern hardware. Consequently, micro-architectural behavior of modern OLAP systems remains unclear. This work presents a micro-architectural analysis of a set of OLAP systems. The results show that traditional commercial OLAP systems suffer from their long instruction footprint, which results in high response times. High-performance columnstores execute tight instruction streams; however, they spend 25 to 82% of their CPU cycles on stalls both for sequential- and random-access-heavy workloads. Concurrent query execution can improve the utilization, but it creates interference in the shared resources, which results in sub-optimal performance.

Micro-architectural analysis is critical to investigate the interaction between workloads and processors. While today's aggressive out-of-order processors provide a rich set of performance events for deep execution cycle analysis, OLTP characterization studies usually use a cache-miss-based method (CMBM). In this work, we investigate the validity and the functionality of CMBM by comparing it with Intel's state-of-the-art Top-down Micro-architecture Analysis Method (TMAM) for OLTP workloads. We show that, while CMBM and TMAM provide a similar high-level micro-architectural behavior, it is inadequate for a fine-grained micro-architectural analysis. We further show that TMAM underestimates memory stalls. We optimize TMAM's execution cycle breakdown, and improve its estimation of memory stalls up to 50%.

Although scaling out of low-power cores is an alternative to power-hungry Intel Xeon processors for reducing the power overheads, they have proven inadequate for complex, non-parallelizable workloads. On the other hand, by the introduction of the 64-bit ARMv8 architecture, traditionally low power ARM processors have become powerful enough to run computationally intensive server-class applications.

Micro-architectural behavior of traditional disk-based online transaction processing (OLTP) systems has been investigated extensively over thepast couple of decades. Results show that traditional OLTP mostly under-utilize the available micro-architectural resources. In-memory OLTP systems, on the other hand, process all the data in main-memory, and therefore, can omit the buffer pool. In addition, they usually adopt more lightweight concurrency control mechanisms, cache-conscious data structures, and cleaner codebases since they are usually designed from scratch. Hence, we expect significant differences in micro-architectural behavior when running OLTP on platforms optimized for in-memory processing as opposed to disk-based database systems. In particular, we expect that in-memory systems exploit micro architectural features such as instruction and data caches significantly better than disk-based systems. This paper sheds light on the micro-architectural behavior of in-memory database systems by analyzing and contrasting it to the behavior of disk-based systems when running OLTP workloads. The results show that despite all the design changes, in-memory OLTP exhibits very similar micro-architectural behavior to disk-based OLTP systems: more than half of the execution time goes to memory stalls where L1 instruction misses and the long-latency data misses from the last-level cache are the dominant factors in the overall stall time. Even though aggressive compilation optimizations can almost eliminate instruction misses, the reduction in instruction stalls amplifies the impact of last-level cache data misses. As a result, the number of instructions retired per cycle barely reaches one on machines that are able to retire up to four for both traditional disk-based and new generation in-memory OLTP.