HYBRID COMPUTING SYSTEM

• US 20140047156A1
• Filed: 08/09/2012
• Published: 02/13/2014
• Est. Priority Date: 08/09/2012
• Status: Abandoned Application

• Current Assignee: Agora Tech Developments Ltd.
• Original Assignee: Agora Tech Developments Ltd.
• Inventors: BILLI, Emillo
• Application Number:
  US13/570,352
• Publication Number:
  US 20140047156A1
• Time in Patent Office: Days
• US Class Current: 710/314
• CPC Class Codes: G06F 13/4022 using switching circuits, e.g. switching matrix, connection or expansion network G06F13/4009 takes precedence
• CPC Class Codes: G06F 2213/0026 PCI express

Abstract

A hybrid computing system comprising: a network fabric; at least one Root Complex board (RCB) and at least one Endpoint Board (EB). Each Root Complex board (RCB) comprises a first processor; a PCIe root complex connected to the first processor; and a first PCIe network switch directly connected to the PCIe root complex. Each Endpoint Board (EB) comprises a second processor; a PCIe interface connected to the second processor; and a second PCIe network switch connected to the PCIe interface. The PCIe network switches of each board (RCB, EB) are connected to the network fabric wherein each Root Complex board (RCB) and each Endpoint Board (EB) are configured for simultaneous use within the hybrid computing system.

Claims

1. A hybrid computing system (100) comprising:

a network fabric (8)

at least one Root Complex board (RCB) comprising;

a first processor (1);

a PCIe root complex (2) connected to the first processor (1);

a first PCIe network switch (4) directly connected to the PCIe root complex (2);

at least one Endpoint Board (EB) comprising;

a second processor (10);

a PCIe interface (11) connected to the second processor (10);

a second PCIe network switch (16) connected to the PCIe interface (11);

wherein;

the first PCIe network switch (4) of each Root Complex board (RCB) is connected to the network fabric (8),

the second PCIe network switch (16) of each Endpoint Board (EB) is connected to the network fabric (8),

each Root Complex board (RCB) and each Endpoint Board (EB) are configured for simultaneous use within the hybrid computing system (100).

2. A hybrid computing system (100) according to claim 1

wherein:

a first dedicated port (3a) of the first PCIe network switch (4) is directly connected to the PCIe root complex (2) within the respective Root Complex board (RCB) through a first main link (3); and

a second dedicated port (12a) of the second PCIe network switch (16) is directly connected to the PCIe interface (11) within the respective Endpoint Board (EB) through a second main link (12).

3. A hybrid computing system (100) according to claim 2

wherein:

the first dedicated port (3a) and the first main link (3) are configured to perform communication with the first processor (1) via the root complex (2) using non-transparent shared memory operations; and

the second dedicated port (12a) and the second main link (12) are configured to perform communication with the second processor (10) via the PCIe interface (11) using direct-shared memory operations.

4. A hybrid computing system (100) according to claim 1

wherein:

the PCIe network switch (4, 16) of each board (RCB, EB) is connected to the network fabric (8) through at least one switch port (5a) of the respective PCIe network switch (4, 16) and at least one corresponding fabric port (8a_1-n) of the network fabric (8) via a corresponding fabric link (5), said fabric link(s) (5) being configured to perform non transparent shared memory operations via the network fabric (8) for inter nodes communications between any Root Complex board (RCB) and/or Endpoint Board (EB).

5. A hybrid computing system (100) according to claim 4,

wherein it further comprises a backplane (BK) for receiving said fabric ports (8a_1-n) corresponding to each board (RCB, EB), wherein the backplane (BK) comprises:

n backplane-to-board connectors (20) corresponding to each of said fabric ports (8a_1-n), wherein each backplane-to-board connectors (20) comprises;

a PCI connector (21) used to connect the boards (RCB, EB) to the network fabric (8), and

power connectors (23) for supplying the boards (RCB, EB) with electrical power;

n pairs of brackets (24, 25) to support the boards (RCB, EB) and to aid in correctly plugging the boards (RCB, EB) into the backplane-to-board connectors (20).

6. A hybrid computing system (100) according to claim 1

wherein:

at least one IO-Fabric (9) is provided within at least one board (RCB, EB);

said IO-fabric (9) is connected to the PCIe network switch (4, 16) of the corresponding board (RCB, EB) via an IO-port (6a) of the PCIe network switch (4, 16) by a corresponding IO-link (6), said IO-Link (6) being configured for direct communication between the processor (1, 10) and IO-Fabric (9),

said IO-port (6a) is configured to perform transparent shared memory communications;

the PCIe network switches (4, 16) of the corresponding boards (RCB, EB), the network fabric (8) and the at least one IO-fabric (9) are configured such that each processor (1, 10) of each board (RCB, EB) is connectable to the at least one IO-Fabric (9) using remote PCIe memory mapping.

7. A hybrid computing system (100) according to claim 1

wherein the network fabric (8) comprises standard PCIe switches, preferably switches compliant with specification of PCI-SIG PCIw gen 2.1.

8. A hybrid computing system (100) according to claim 1

wherein the first processor (1) and/or the second processor (10) is one of the following:

a central processing unit (CPU) such as a microprocessor;

a Field-programmable gate array (FPGA);

an accelerator, such as a graphics processing unit (GPU);

a Digital signal possessor (DSP);

a physics processing unit (PPU);

a front end processor,

a memory controller;

a drive controller such as solid-state drive (SSD) controller, a hard drive-controller or a disk array controller;

or

a network interface controller (NIC), such as an InfiniBand controller.

9. A hybrid computing system (100) according to claim 1

wherein the network fabric (8) is a PCIe cluster of one of the following network topologies:

point-to-point;

bus;

star;

ring, preferably a 3D torus;

mesh;

or

tree.

10. A hybrid computing system (100) according to claim 1

wherein it is configured to receive a number of n (up to 32) boards (1 or 10), the network fabric (8) comprising n fabric ports (8a_1-n) for receiving any combination of Root Complex boards (RCB) and/or Endpoint Boards (EB) configured for simultaneous use within the hybrid computing system (100).