Abstract: Heat dissipation system, a power converter using such a heat dissipation system, and an associated method of thermal management of the power converter are disclosed. The heat dissipation system includes a condenser, a first cooling loop, and a second cooling loop. The first cooling loop is coupled to the condenser and includes a first two-phase heat transfer device. The second cooling loop is coupled to the condenser and includes a second two-phase heat transfer device. The condenser is disposed above the first and second two-phase heat transfer devices.

Abstract: A primary storage array and a remote storage array each create compressibility heat maps and periodically exchange compressibility heat maps over a remote replication facility. The primary R1 storage array also forwards an Activity Based Compression (ABC) heat map describing extent IO activity to the remote storage array. The primary and remote storage arrays use the compressibility heat maps to update per-extent compressibility forecast models to forecast compressibility values for extents during an upcoming time interval. The storage arrays use both activity-based scores and forecast compressibility scores to select extents to be stored in compressed and uncompressed form.

Abstract: A snapshot resource manager running on a storage array forecasts the time-variable need for storage array resources for generation of snapshots. The forecasted resources are accumulated in pools that are pre-allocated for snapshot generation such that the required resources are available when receipt of commands to generate snapshots are received, thereby mitigating the performance impact of snapshot generation on servicing of IOs from instances of host applications.

Abstract: In a data storage system that includes two or more storage engines in interconnected by a channel-based fabric, highly deduplicated data is identified and promoted to a dedupe performance tier characterized by mirrors on each of the storage engines. Blocks of deduplication data are scored based on reference counts and read-miss access patterns. Promotion of blocks to the dedupe performance tier and demotion of blocks from the dedupe performance tier are calculated based on the scores. Performance is enhanced by enabling highly deduplicated blocks to be accessed from engine-local memory and engine-local storage via switches rather than the channel-based fabric.

Abstract: Tiers of compression algorithms may be determined using compression information collected regarding compression ratios achieved for data sets using compression algorithms. Each tier may meet specified criteria regarding expected compression ratios achieved for a specified portion or number of data sets. Compression algorithms of each tier may be implemented by a different hardware device that may include hardware accelerators for the algorithms of the tier. Different tiers, and thus different hardware devices, achieve different levels of compression. A recommendation may be provided using compression information collected, such as from one of the hosts, regarding which hardware device to use for compression. The recommendation may be to purchase a license to use or whether to purchase a particular hardware device for compression. Compression information may be collected by a host that issues tagged I/Os providing a hint regarding what compression algorithm to use for the particular I/O operation data.

Abstract: Streaming machine telemetry (SMT) event counters are placed in critical code paths of software executing on a storage system. Each monitoring interval the values of the SMT counters are reported. When a critical error occurs on the storage system, a time series set of SMT counters from a set of previous monitoring intervals is labeled with the error type and used as a training example for a learning process. The learning process is trained to learn to learn recursions between time series sets of SMT counter values and labeled error types. Once trained, a checkpoint of the learning model is deployed as an inference model and used to predict the likely occurrence of errors before the errors occur. Predicted errors are logged into a proactive service request queue, and feedback related to the predicted errors are used as feedback to continue training the learning process.

Abstract: Techniques are disclosed relating to determining a similarity of components of a current webpage to different UI components for use in automatically generating an updated webpage. A computer system may receive information specifying a current webpage, including a particular current UI component and information specifying a plurality of different UI components for an updated webpage. The computer system may identify one or more characteristics of the particular current UI component. The computer system may determine, based on the identified one or more characteristics, a similarity of ones of the plurality of different UI components to the particular current UI component. The computer system may select, based on the determining, a particular different UI component from the plurality of different UI components for use, in the updated webpage, for the particular current UI component. Such techniques may advantageously improve user experience by automatically providing up-to-date user interfaces.

Abstract: Domain-specific initial high-speed memory configuration policies are created by analyzing workloads of storage systems deployed in each of a plurality of customer domains. Each domain-specific initial high-speed memory configuration policy specifies the size of both a mirrored policy region and a non-mirrored policy region. The sizes of the mirrored and non-mirrored policy regions are based on the respective proportion of read IO and write IO operations in the analyzed workloads of each customer domain. Each domain-specific initial high-speed memory configuration policy also specifies a distribution of slot sizes that should be created in each of the policy regions, based on read IO and write IO histograms created from the analyzed workloads in each customer domain. When a new storage system is deployed, the customer domain where the new storage system will be used is determined, and the high-speed memory configuration policy for the domain is applied to the storage system.

Abstract: A data storage system includes a tuner that obtains data samples for data storage operations of workloads and calculates feature measures for a set of features of the data storage operations over aggregation intervals of an operating period. It further (1) applies a cluster analysis to the feature measures to define a set of clusters, and assigns the feature measures to the clusters, and (2) applies a classification analysis to the feature measures labelled by their clusters to identify dominating features of each cluster, and generates workload profiles for the clusters based on the dominating features, and then automatically adjusts configurable processing mechanisms (e.g., caching or tiering) based on the workload profiles and performance or efficiency goals.

Abstract: A system can train a generator neural network to produce a trained generator neural network of a generative adversarial network, wherein the trained generator neural network is configured to output a bill of materials in response to receiving functional requirements for a computer system. The system can train a discriminator neural network to produce a trained discriminator neural network of the generative adversarial network, wherein the trained discriminator neural network is configured to output whether the bill of materials received from the trained generator neural network satisfies the functional requirements for the computer system. The system can produce an output bill of materials from the generative adversarial network based on the functional requirements. The system can store the output bill of materials in the system.

Abstract: Feature usage profiles of shared global memory are monitored and the profiles are used to allocate computational resources (CPU cores) between thread groups. A global memory monitor is used to monitor global memory metadata to determine the amount of space currently being consumed by a particular feature or service, and dynamically adjusts the amount of computational resources allocated to the thread group responsible for processing IO operations related to the feature or service. By exploiting an inversely proportional relationship between allocated computational resources and usage of shared global memory, it is possible to monitor feature usage profiles of shared global memory to dynamically allocate computational resources to thread groups. By looking at anomalies in usage of slots of shared global memory, and in a targeted way speeding up the processes working on such slots, it is possible to improve overall utilization of shared global memory.

Abstract: Embodiments of the present disclosure relate to establishing persistent cache memory as a write tier. An input/output (IO) workload of a storage array can be analyzed. One or more write data portions of the IO workload can be stored in a persistent memory region of one or more disks of the storage array.

Abstract: Non-volatile Random Access Memory (NVR) on a storage system may be dynamically converted between use as temporary memory in a memory context and use as persistent memory in a storage context. NVR (e.g., embodied as DIMM) may be utilized in a hybrid capacity, where some of the NVR is used as memory and some of the NVR is used as storage, and where NVR memory is converted to memory as needed, dynamically as I/O is being processed using the NVR. A host system may be directly connected to an internal switching fabric of the data storage system without an intervening component of the storage system (e.g., a director) controlling access of the host system to the internal fabric or to the memory. The host system may provision and use the NVR as storage by directly communicating with the NVR over the internal fabric, for example, using RDMA.

Abstract: Different storage platforms and different applications generate different types of traffic. Separate learning processes are trained using live IO operations from different types of reference workloads. The reference workloads are based on multiple types of applications executing in multiple types of environments. Example applications may include Oracle, SAP, SQL, etc. Example environments may include finance/banking, retail, on-line transaction processing, healthcare, etc. Once the learning processes have been trained, trained model checkpoints describing the trained learning processes are stored in a reference workload repository. The trained model checkpoints are used in a testing environment to enable a test server to create a test model configured to generate simulated test IO operations to be applied to a storage system. The generated test IO operations simulate the workload from an intended application and environment where the storage system will be used.

Abstract: The over-provisioning (OP) of a physical storage device (PSD) may be increased, and the useful life of the PSD increased, by converting uncompressed data stored on the PSD to compressed data. It may be determined that increasing the useful life of the PSD, and the data reduction resulting from the compression, outweigh the benefit of faster I/O response times if the data remains uncompressed. A first portion of the PSD may be initially reserved for compression. A second portion of the PSD may store compressed data. It may be determined whether it is desirable to increase the OP of the PSD to thereby reduce the effective write rate on the PSD. If compression is determined to be desirable, the dynamic portion may be compressed, thereby reducing the amount of storage space consumed by the data, and freeing up storage space that can be used by the PSD for OP.

Abstract: A data storage system with interconnected compute nodes includes a shared memory with volatile and non-volatile portions. Data tracks evicted from the volatile portion are moved to the non-volatile portion based on a cache-miss interarrival rate threshold that is calculated based on capacity and fall-through time of the non-volatile portion of the shared memory. Data extents on non-volatile drives are characterized based on dominant modes of extent-level cache-miss interarrival histograms generated using countdown timers of most recent backend accesses of sub-extents. The dominant mode of the extent in which a backend track evicted from the volatile portion of the shared memory resides is compared with the threshold in order to determine whether to move the backend track to the non-volatile portion of the shared memory.

Abstract: The present disclosure relates to one or more memory management techniques. In embodiments, one or more regions of storage class memory (SCM) of a storage array is provisioned as expanded global memory. The one or more regions can correspond to SCM persistent cache memory regions. The storage array's global memory and expanded global memory can be used to execute one or more storage-related services connected to servicing (e.g., executing) an input/output (IO) operation.

Abstract: Embodiments of the present disclosure relate to autonomous data protection. An input/output (I/O) stream can be received for a storage device. One or more anomalies corresponding to the I/O stream can be identified. At least one of the one or more anomalies can be offloaded anomalies to a remote storage based on a capacity of memory allocated to store at least one snapshot of the storage device that include at least one of the one or more anomalies.

Abstract: IO traces on a high-speed memory that provides temporary storage for multiple storage volumes are stored in a trace buffer. IO operations on different storage volume are considered separate workloads on the high-speed memory. Periodically, the IO traces are processed to extract workload features for each workload. The workload features are stored in a feature matrix, and the workload features from multiple IO trace buffer processing operations are aggregated over time. A HDBSCAN unsupervised clustering machine learning process is used to create a set of four workload clusters and an outlier cluster. A dominant feature of each workload cluster is used to set a policy for the workload cluster. IO percentages for clusters with the same policies are used to set minimum sizes for policy regions in the high-speed memory. Histograms based on the workloads are used to determine segmentation rules specifying slot sizes for the policy regions.