Processor-embedded distributed smart disks for I/O-intensive workloads: architectures, performance models and evaluation

Network and device-level impacts: performance and reliability of active I/O storage systems

“…The storage optimization techniques put forth in smart disks [4] are also applicable here, since the storage is also programmed in software using general-purpose processors. Thus, processing power at the storage level is readily available for applicationlevel functions as well as storage optimizations.…”

“…To optimize the performance of these queries on an active storage system, the work in [7] proposed the "on-device operations bundling" method to combine elementary database operations that may share common data dependencies. The work in [3] explained the processing model for the TPC-H queries, which resulted both from the software architecture proposed in [4] and the operations bundling technique. T5: HCA places data onto network; T6: HCA updates descriptor for send completion; R1: post receive doorbell to HCA; R2: HCA moves doorbell to receive doorbell queue; R3: HCA receives inbound data into receive staging area; R4: HCA transfers receive descriptor to staging area; R5: HCA transfers inbound data to receive buffer; and R6: HCA updates descriptor for receive completion.…”

Fault Recovery Designs for Processor-Embedded Distributed Storage Architectures with I/O-Intensive DB Workloads

“…Our previous work [2,3] proposed distributed disk-and MEMS-based smart storage systems, and provided the performance and processing models to evaluate their potential benefits. Such a system takes advantage of the increasing processing power on the storage device by performing I/O-intensive operations directly at the device and in parallel across multiple devices.…”

“…For distributed storage architectures, operations bundling is a logical approach to improve performance. With software architecture specified in [2], processing model of a smart storage system will be as illustrated in Figure 3, using disk-based systems as an example. An application inputs a query or executable code through the parser that decides the input as either valid (and passes it to the optimizer) or invalid (and passes it to the error handler).…”

“…Upon completion, the system responds to the user with the final results through the designated SNs. The software architecture proposed in [2] provides an ondevice kernel (SD OS) to support the execution control, data access, memory buffer management, signaling, code/data recovery, etc. to facilitate this processing model.…”