Dynamic memory for compute resources in a data center
Examples may include sleds for a rack in a data center including physical compute resources and memory for the physical compute resources. The memory can be disaggregated, or organized into first level and second level memory. A first sled can comprise the physical compute resources and a first set of physical memory resources while a second sled can comprise a second set of physical memory resources. The first set of physical memory resources can be coupled to the physical compute resources via a local interface while the second set of physical memory resources can be coupled to the physical compute resources via a fabric.
1. A system for a data center comprising:
a rack comprising a plurality of sled spaces;
at least two or more compute sleds coupled to the rack via at least two or more of the plurality of sled spaces, each of the two or more compute sleds comprising:
at least one physical compute resource;
a set of physical first level memory resources coupled to the at least one physical compute resource;
a first compute sled memory controller to couple the at least one physical compute resource to the set of physical first level memory resources;
a second compute sled memory controller to couple the at least one physical compute resource to a portion of a set of physical second level memory resources; and
a compute sled fabric interface to couple to a fabric; and
a memory resource sled coupled to the rack via another one of the plurality of sled spaces different than the two or more of the plurality of sled spaces, the memory resource sled comprising:
the set of physical second level memory resources; and
a second level memory sled fabric interface to couple, to each of the at least one physical compute resources of the two or more compute sleds via the fabric and the compute sled fabric interfaces, a portion of the set of physical second level memory resources, the portion of the set of physical second level memory resources coupled to a first one of the at least two or more compute sleds different than the portion of the set of physical second level memory resources coupled to a second one of the at least two or more compute sleds.
2. The system of claim 1 , the sets of physical first level memory resources comprising volatile memory.
3. The system of claim 2 , the set of physical second level memory resources comprising byte-addressable write-in place nonvolatile memory.
4. The system of claim 1 , each of the at least one physical compute resources comprising a central processing unit, a field programmable gate array, a graphics processing unit, or an application specific integrated circuit.
5. The system of claim 1 , the at least one physical compute resource of each of the two or more compute sleds coupled to the set of physical first level memory resources via a local interface, the local interface a peripheral component interconnect express compliant standard interface.
6. The system of claim 1 , the fabric an optical fabric.
7. The system of claim 1 , comprising a virtual infrastructure management framework to dynamically allocate the portions of the physical second level memory resources to each of the two or more compute sleds.
8. The system of claim 1 , wherein the portion of the set of physical second level memory resources coupled to the first one of the at least two or more compute sleds to be dynamically adjusted based on a workload to be executed on the at least one physical compute resource of the first one of the at least two or more compute sleds.
9. An apparatus for a compute resource sled of a data center, comprising:
at least one physical compute resource;
a set of physical first level memory resources coupled to the at least one physical compute resource;
a first compute sled memory controller to couple the at least one physical compute resource to the set of physical first level memory resources;
a fabric interface, the fabric interface to couple the at least one physical compute resource to a portion of a set of physical second level memory resources via a fabric, the set of physical second level memory resources to be shared between the compute resource sled and another compute resource sled of the data center; and
a second compute sled memory controller to couple the at least one physical compute resource to the portion of the set of physical second level memory resources, the portion of the set of physical second level memory resources coupled to the compute resource sled of the data center different than the portion of the set of physical second level memory resources coupled to the other compute resource sled of the data center.
10. The apparatus of claim 9 , the first set of physical memory resources comprising volatile memory.
11. The apparatus of claim 10 , the second set of physical memory resources comprising three-dimensional (3D) cross-point memory.
12. The apparatus of claim 9 , the at least one physical compute resource comprising a central processing unit, a field programmable gate array, a graphics processing unit, or an application specific integrated circuit.
13. The apparatus of claim 9 , the at least one physical compute resource coupled to the set of physical first level memory resources via a local interface, the local interface a peripheral component interconnect express compliant standard interface.
14. The apparatus of claim 9 , the fabric an optical fabric.
15. The apparatus of claim 9 , wherein the portion of the set of physical second level memory resources coupled to the compute resource sled of the data center to be dynamically adjusted based on a workload to be executed on the at least one physical compute resource of the compute resource sled of the data center.
16. A method comprising:
allocating a first portion of a set of physical second level memory resources to a first one of a plurality of compute sleds;
allocating a second portion of the set of physical second level memory resources, different than the first portion of the set of physical second level memory resources, to a second one of the plurality of compute sleds;
coupling, via a fabric, at least one physical compute resource of the first one of the plurality of compute sleds to the first portion of the set of physical second level memory resources; and
coupling, via the fabric, at least one physical compute resource of the second one of the plurality of compute sleds to the second portion of the set of physical second level memory resources, wherein each compute sled of the plurality of compute sleds comprise: a first compute sled memory controller to couple the at least one physical compute resource of the compute sled to a set of physical first level memory resources of the compute sled, a fabric interface to couple the at least one physical compute resource of the compute sled to the respective portion of the set of physical second level memory resources via the optical fabric, and a second compute sled memory controller to couple the at least one physical compute resource of the compute sled to the respective portion of the set of physical second level memory resources.
17. The method of claim 16 , the set of physical second level memory resources comprising three-dimensional (3D) cross-point memory.
18. The method of claim 16 , the at least one physical compute resource comprising a central processing unit, a field programmable gate array, a graphics processing unit, or an application specific integrated circuit.
19. The method of claim 16 , comprising:
composing a first virtual compute platform from the at least one physical compute resource of the first compute sled and the first portion of the set of physical second level memory resources; and
allocating the first portion of the set of physical second level memory resources to the first compute sled based on a workload to be executed by the first virtual compute platform.
20. The method of claim 19 , comprising:
composing a second virtual compute platform from the at least one physical compute resource of the second compute sled and the second portion of the set of physical second level memory resources; and
allocating the second portion of the set of physical second level memory resources to the second compute sled based on a workload to be executed by the second virtual compute platform.
21. The method of claim 20 , comprising adjusting the allocation of at least one of the first portion of the set of physical second level memory resources or the second portion of the set of physical second level memory resources.
22. The method of claim 16 , each of the first compute sled and the second compute sled comprising a set of physical first level memory resources, allocating the first portion of the set of physical second level memory resources to the first compute sled based on a quantity of the set of physical first level memory resources of the first compute sled and allocating the second portion of the set of physical second level memory resources to the second compute sled based on a quantity of the set of physical first level memory resources of the second compute sled.
23. The method of claim 22 , comprising allocating the portions of the set of physical second level memory resources to the first and the second compute sleds to achieve a ratio of the quantity to the sets of physical first level memory resources to a quantity of the portions of the set of physical second level memory resources.
24. The method of claim 16 , comprising dynamically adjusting the first portion of the set of physical second level memory resources based on a workload to be executed on the at least one physical compute resource of the first one of the plurality of compute sleds.