Alleviating interconnect traffic in a disaggregated memory system
One or both of read and write accesses to a fabric-attached memory module via a fabric interconnect are monitored. In one or more implementations, offloading of one or more tasks accessing the fabric-attached memory module to a processor of a routing system associated with the fabric-attached memory module is initiated based on the read and write accesses to the fabric-attached memory module. Additionally or alternatively, replicating memory of the fabric-attached memory module to a cache memory of a computing node in the disaggregated memory system executing one or more tasks of a host application is initiated based on the write accesses to the fabric-attached memory module.
1 . A system comprising:
a processor of a computing node configured to:
send read and write accesses to a fabric-attached memory via a fabric interconnect of a disaggregated memory system; and
transfer, based on the read and write accesses, one or more tasks accessing the fabric-attached memory to a processor of a routing system associated with the fabric-attached memory and at least one additional fabric-attached memory via the fabric interconnect.
2 . The system of claim 1 , wherein the transfer is further based on a number of tasks accessing the fabric-attached memory, a total read traffic to the fabric-attached memory, a total write traffic to the fabric-attached memory, a number of the read and write accesses to the fabric-attached memory, or a frequency of the read and write accesses to the fabric-attached memory.
3 . The system of claim 1 , wherein the transfer is initiated in response to an indication that the read and write accesses satisfy a rule.
4 . The system of claim 3 , wherein the rule includes a total read and write traffic to the fabric-attached memory over a duration of time exceeding a threshold amount, the threshold amount being a percentage of a total read and write traffic to the fabric-attached memory and at least one additional fabric-attached memory coupled to the routing system.
5 . The system of claim 4 , wherein the duration of time is a configurable parameter.
6 . The system of claim 4 , wherein the duration of time is a rolling window.
7 . The system of claim 3 , wherein the rule includes a total read and write traffic to the fabric-attached memory and at least one additional fabric-attached memory coupled to the routing system over a duration of time exceeding a threshold amount, the threshold amount being a percentage of a total read and write traffic to multiple fabric-attached memories coupled to the routing system.
8 . The system of claim 3 , wherein the transfer is ceased in response to an indication that the read and write accesses no longer satisfy the rule.
9 . The system of claim 8 , wherein the rule includes:
a first threshold amount to determine that the read and write accesses satisfy the rule; and
a second threshold amount to determine that the read and write accesses no longer satisfy the rule, the second threshold amount being different than the first threshold amount.
10 . The system of claim 1 , wherein the one or more tasks include one or more threads of a host application running in the disaggregated memory system.
11 . The system of claim 1 , wherein the one or more tasks include each task of a host application running in the disaggregated memory system that read from or write to the fabric-attached memory.
12 . The system of claim 1 , wherein the processor is further configured to receive, based on write accesses to the fabric-attached memory, a replicated memory of the fabric-attached memory to a cache memory of the computing node.
13 . A system comprising:
a router configured to:
receive, from a processor of a computing node, read and write accesses to a fabric-attached memory via a fabric interconnect of a disaggregated memory system; and
receive, based on the read and write accesses, an offloading of one or more tasks accessing the fabric-attached memory from the processor of the computing node to an additional processor communicatively coupled to the router within a routing system, the additional processor communicatively coupled to the fabric-attached memory and at least one additional fabric-attached memory via the fabric interconnect.
14 . The system of claim 13 , wherein the offloading is further based on a number of tasks accessing the fabric-attached memory, a total read traffic to the fabric-attached memory, a total write traffic to the fabric-attached memory, a number of the read and write accesses to the fabric-attached memory, or a frequency of the read and write accesses to the fabric-attached memory.
15 . The system of claim 13 , wherein the offloading is initiated in response to a determination that the read and write accesses satisfy a rule.
16 . The system of claim 15 , wherein the rule includes a total read and write traffic to the fabric-attached memory and at least one additional fabric-attached memory coupled to the routing system over a duration of time exceeding a threshold amount, the threshold amount being a percentage of a total read and write traffic to multiple fabric-attached memories coupled to the routing system.
17 . The system of claim 15 , wherein the offloading is ceased in response to an indication that the read and write accesses no longer satisfy the rule.
18 . The system of claim 17 , wherein the rule includes:
a first threshold amount to determine that the read and write accesses satisfy the rule; and
a second threshold amount to determine that the read and write accesses no longer satisfy the rule, the second threshold amount being different than the first threshold amount.
19 . A method comprising:
receiving, based on read and write accesses to a fabric-attached memory via a fabric interconnect of a disaggregated memory system, an indication to offload one or more tasks; and
transferring, in response to the indication, one or more tasks accessing the fabric-attached memory from a processor of a computing node to an additional processor of a routing system associated with the fabric-attached memory via the fabric interconnect, the additional processor communicatively coupled to the fabric-attached memory and at least one additional fabric-attached memory.
20 . The method of claim 19 , wherein the indication is further based on a number of tasks accessing the fabric-attached memory, a total read traffic to the fabric-attached memory, a total write traffic to the fabric-attached memory, a number of the read and write accesses to the fabric-attached memory, or a frequency of the read and write accesses to the fabric-attached memory.