Exactly-once transaction semantics for fault tolerant fpga based transaction systems

A Field Programmable Gate Array (FPGA) based transaction system (200) comprising:

• a plurality of FPGAs including an active FPGA and remaining passive FPGAs, each of the plurality of FPGAs being configured to cooperate with at least one client via a Transmission Control Protocol (TCP) connection, each of the plurality of FPGAs having one or more storage devices being a combination of a volatile memory and a non-volatile memory and implemented as a key-value store, the plurality of FPGAs comprising:
• a Modified Transmission Control Protocol Unit (MTCPU) configured to:
• receive TCP packets corresponding to incoming requests, by the MTCPU of the active FPGA, from the at least one client;
• append the TCP packets, with at least a time-stamp as a non-deterministic parameter;
• publish the TCP packets, by the MTCPU of the active FPGA, to the passive FPGAs, and thereafter transmit a TCP acknowledgement of the incoming requests to the at least one client;
• control, by the MTCPU of the active FPGA, scheduling of the incoming requests along with resubmitted requests caused by concurrent execution by an application after a configurable delay, wherein the resubmitted requests are prioritized over the incoming requests;
• and
• transmitting (a) a response or (b) the response and one or more notifications to the at least one client;
• a Message Parsing Unit (MPU) configured to:
• detect, by the MPU of the active FPGA, duplicate requests received based on associated request identifiers;
• retrieve, by the MPU of the active FPGA, the resubmitted requests based on an associated client identifier and an associated request identifier;
• and
• parse and convert the TCP packets received in a sequence from the MTCPU into a message format associated with a network between the at least one client and the plurality of FPGAs;
• a Transaction Processing Unit (TPU) configured to:
• host the application which processes the incoming requests or the resubmitted requests from the MPU and generates (a) the response or (b) the response and the one or more notifications;
• and
• receive the resubmitted requests caused by the concurrent execution from the application and send to the MTCPU for scheduling via the MPU;
• a Transaction Commit Unit (TCU) configured to:
• receive (a) the response or (b) the response and the one or more notifications sent from the TPU;
• generate messages comprising the incoming requests or the resubmitted requests and (a) the response or (b) the response and the one or more notifications, wherein the TCP packets constitute (a) at least a portion of a message or (b) one or more of the messages;
• and
• transmit the messages to the MTCPU;
• and
• a Monitoring Unit (MU) configured to:
• monitor and initiate an FPGA failover process in the event an active FPGA failure is detected.

A method (300) comprising:

• establishing a Transmission Control Protocol (TCP) connection from at least one client to each Field Programmable Gate Array (FPGA) and each Host in an FPGA based transaction system comprising a plurality of FPGAs, a plurality of Hosts and at least one client with messages transmitted therebetween, wherein each of the messages comprise an incoming request from the at least one client and (a) a response or (b) the response and one or more notifications associated thereof, to the at least one client, and wherein an incoming request and a corresponding response represents a transaction (302);
• switching a mode of operation of the FPGA based transaction system between a high performance mode of operation and a conservative mode of operation based on a pre-configured mode switching threshold depending on number of available Hosts in the plurality of Hosts, wherein (i) the high performance mode involves saving the messages in a volatile memory associated with each of the plurality of Hosts and thereafter transmitting an internal acknowledgement thereof to a corresponding FPGA in the plurality of FPGAs, and (ii) the conservative mode involves saving the messages in a non-volatile memory associated with each of the plurality of Hosts and thereafter transmitting an internal acknowledgement thereof to a corresponding FPGA in the plurality of FPGAs (304);
• saving the messages from the volatile memory into the non-volatile memory of the plurality of Hosts during a switching period when the high performance mode of operation switches to a conservative mode of operation (306);
• delaying transmitting of (a) the response or (b) the response and the one or more notifications to the at least one client, till the end of saving of the associated messages into the non-volatile memory during the switching period or based on a pre-configured time delay to ensure an incoming request is received by the plurality of Hosts instead of an associated duplicate request that reaches the plurality of Hosts prior to the incoming request (308);
• and
• delaying transmitting of the internal acknowledgement to the plurality of FPGAs, till the end of saving of the associated messages into the non-volatile memory (310).