US 12,001,352 B1Granted

Transaction ordering based on target address

AssigneeAmazon Technologies, Inc.

InventorsRashika Kheria, Ron Diamant, Se Wang Oh, Guy Nakibly

Life of this patent

Estimating the term…

Loading patent…

Abstract

Techniques are provided to maintain data coherency for data transfers among data processing devices in a distributed computing environment. A data buffer in each data processing device can be mapped to an address range that is assigned to transactions that allow out-of-order completions, and a message buffer in each data processing device can be mapped to an address range that is assigned to transactions that follow transaction ordering. Thus, a transaction to store a set of data into the data buffer is completed before a transaction to write a synchronization message in the message buffer indicating that the set of data is stored in the data buffer based on the mapping irrespective of the transaction ordering indicated by each transaction.

Claims20

Dependent claims

A computing system comprising:

a host device comprising a host processor;
a network device coupled to the host device via an interconnect, the network device comprising a network interface to communicate with a remote computing system via a network;
and
an accelerator device coupled to the host device and the network device via the interconnect, the accelerator device comprising a data buffer and a message buffer, wherein the data buffer is mapped to a first address range and the message buffer is mapped to a second address range, and wherein the first address range is assigned to transactions that follow a relaxed-order completion and the second address range is assigned to transactions that follow a strong-order completion,
wherein the network device is configured to:
send, via the interconnect, a set of first transactions to store a set of data into the data buffer of the accelerator device without going through the host device;
and
send, via the interconnect, a second transaction with a write completion message to the host device indicating that the set of data has been sent to the accelerator device,
wherein the host device is configured to:
upon receiving the second transaction, send, via the interconnect, a third transaction with a synchronization message to the accelerator device to initiate computations using the set of data, and
wherein the accelerator device is configured to:
receive, via the interconnect, the set of first transactions from the network device;
receive, via the interconnect, the third transaction from the host device;
treat the set of first transactions as transactions that follow the relaxed-order completion based on mapping of the data buffer to the first address range, and treat the third transaction as a transaction that follows the strong-order completion based on the mapping of the message buffer to the second address range;
and
complete the set of first transactions before executing the third transaction by storing the set of data into the data buffer before storing the synchronization message in the message buffer to ensure that a complete set of data is available in the data buffer to be used for the computations.

13 more elements

A method performed by a first integrated circuit (IC) device, comprising:

receiving, via an interconnect, a set of first transactions from a second IC device to store a set of data into a data buffer that is mapped to a first address range of the first IC device, the first address range being assigned to transactions that allow out-of-order completion, wherein the second IC device sends a second transaction with a write completion message to a host processor after sending the set of first transactions to the first IC device;
receiving, via the interconnect, a third transaction to write a synchronization message to a message buffer issued by the host processor upon receiving the second transaction, wherein the message buffer is mapped to a second address range of the first IC device, the second address range being assigned to transactions that follow transaction ordering;
and
completing the set of first transactions before executing the third transaction in the first IC device by storing the set of data into the data buffer before allowing the synchronization message to be written to the message buffer based on the mapping of the data buffer to the first address range and the mapping of the message buffer to the second address range.

1 more element

A first integrated circuit (IC) device comprising:

a data buffer to store data for computations, the data buffer being mapped to a first address range of the first IC device that allows transactions to have out-of-order completion;
and
a message buffer to store messages, the message buffer being mapped to a second address range of the first IC device that requires transactions to follow transaction ordering,
wherein the first IC device is configured to:
receive, via an interconnect, a set of first transactions from a second IC device to store a set of data into the data buffer, wherein the second IC device sends a second transaction with a write completion message to a host processor after sending the set of first transactions to the first IC device;
receive, via the interconnect, a third transaction to write a synchronization message to the message buffer issued by the host processor upon receiving the second transaction;
and
complete the set of first transactions before executing the third transaction by storing the set of data into the data buffer before allowing the synchronization message to be written to the message buffer based on the mapping of the data buffer to the first address range and the mapping of the message buffer to the second address range.

5 more elements

Dependencies are inferred from the claim text.