Method for executing queries on data chunks using graphic processing units
The present invention discloses a method for optimizing the throughput of hardware accelerators (HWAs) in a computerized abstraction system, by utilizing the maximal data input bandwidth to the said HWAs. The method is comprised of the following steps: dynamically obtaining the quantities and properties of HWAs and storage units within the computerized abstraction system dynamically allocating cache memory space per each of the HWAs, according to the said obtained quantities and properties, to minimize the time required for reading data from storage instances to the said HWA dynamically allocating spoolers per each of the HWAs, according to the said obtained quantities and properties, to buffer the input data and ensure a continuous flow of input data, in the target HWA's maximal input bandwidth.
1. A method for optimizing the throughput of hardware accelerators (HWAs), by maintaining a maximal rate of data transfer from storage units to the said HWAs, said method comprising the steps of:
storing and managing, by a File-System, access to data on a plurality of storage modules in the computerized abstraction system's environment;
allocating a memory cache space, per each of the HWAs, to minimize the time required for reading data from storage instances to target HWAs;
allocating spoolers, per each of the HWAs, to buffer the input data, and ensure a continuous flow of input data in the target HWA's maximal input bandwidth;
wherein the said memory cache space is optimally allocated, by an Opaque File System, to cache the input data, and minimize the time required for reading data from storage modules to target HWAs;
wherein the said spoolers are optimally allocated, by the Opaque File System, to buffer the input data and ensure a continuous flow of input data, in the target HWA's maximal input bandwidth; and
wherein the allocation of said memory cache space and said spoolers is adapted dynamically according to the current quantities and properties of HWA and storage instances within the computerized system.
2. The method of claim 1 , wherein the said cache memory space is divided between a local memory instance and a remote memory instance, thus forming a hierarchical cache scheme.
3. The method of claim 1 , wherein the spoolers form a data queue at the HWA's input, thus enabling the management of queued input data, for improving the data input rate.
4. The method of claim 1 , wherein:
the HWA can directly retrieve data from adjacent HWAs or storage modules in the said computerized abstraction system, without consuming resources from a common data bus or processor, and
the said retrieved data is cached by the said cache memory, to minimize the time required for reading data from the said adjacent HWAs or storage modules to the target HWA, and
the said retrieved data is buffered by the said data spoolers to ensure a continuous flow of input data in the target HWA's maximal input bandwidth.
5. A computerized abstraction system for optimizing the throughput of hardware accelerators (HWAs), by maintaining a maximal rate of data transfer from storage units to the said HWAs, said system comprised of:
a File-System for storing and managing access to data on a plurality of storage modules in the computerized abstraction systems environment
a memory cache space, allocated per each of the HWAs, to minimize the time required for reading data from storage instances to target HWAs;
spoolers allocated per each of the HWAs, to buffer the input data, and ensure a continuous flow of input data in the target HWA's maximal input bandwidth;
an Opaque File System module that:
optimally allocates the said memory cache space, to cache the input data, and minimize the time required for reading data from storage modules to target HWAs, and
optimally allocate the said spoolers to buffer the input data and ensure a continuous flow of input data, in the target HWA's maximal input bandwidth
wherein the allocation of said memory cache space and said spoolers is adapted dynamically according to the current quantities and properties of HWA and storage instances within the computerized system.
6. The system of claim 5 , wherein the said cache memory space is divided between a local memory instance and a remote memory instance, thus forming a hierarchical cache scheme.
7. The system of claim 5 , wherein the said spoolers form a data queue at the HWA's input, thus enabling the management of queued input data, to improve the data input rate.
8. The system of claim 5 , further comprising a direct memory access module, enabling the HWAs to directly retrieve data from adjacent HWAs or storage modules within the said computerized abstraction system, without consuming resources from a common data bus or processor, wherein
the said retrieved data is cached by the said cache memory, to minimize the time required for reading data from the said adjacent HWAs or storage modules to the target HWA, and
the said retrieved data is buffered by the said data spoolers to ensure a continuous flow of input data in the target HWA's maximal input bandwidth.