Memory migration in presence of live memory traffic

1. A method for migration from a first addressing scheme to a second addressing scheme, comprising:

receiving, by a coherence ordering unit (COU), a first request to access a first memory address and a second request to access a second memory address;

comparing, by the COU, the first memory address and the second memory address with a barrier pointer referencing a barrier address,

wherein the first memory address, the second memory address, and the barrier address are different, and wherein the barrier address separates migrated addresses and un-migrated addresses;

tagging, by the COU, the first request with a first tag indicative of the first addressing scheme in response to the first memory address being on an un-migrated side of the barrier address;

tagging, by the COU, the second request with a second tag indicative of the second addressing scheme in response to the second memory address being on a migrated side of the barrier address; and

sending, by the COU, the first request to a first memory controller unit (MCU) and the second request to a second MCU,

wherein the first MCU interprets the first memory address according to the first addressing scheme in response to the first tag, and

wherein the second MCU interprets the second memory address according to the second addressing scheme in response to the second tag.

2. The method of claim 1 , further comprising:

receiving a third request for write access to the barrier address;

tagging the third request with the second tag;

sending, by the COU, the third request for interpretation after tagging the third request; and

incrementing the barrier pointer to reference a new barrier address in response to the third request,

wherein the migrated addresses are below the new barrier address.

3. The method of claim 1 , further comprising:

receiving, by the COU and from a scrubber application, a read request directed towards the barrier address,

wherein the first request, the second request, and the third request are from a user application;

sending a value stored at the barrier address to the scrubber application in response to the read request;

receiving, by the COU and from the scrubber application, a write request comprising the value and directed toward the barrier address;

tagging, by the COU, the write request with the second tag;

sending the write request for interpretation; and

incrementing the barrier pointer to reference a new barrier address in response to the write request,

wherein the migrated addresses are below the new barrier address.

4. The method of claim 3 , further comprising:

incrementing the barrier pointer to reference the barrier address before the read request directed towards the barrier address is received from the scrubber;

triggering a time window in response to incrementing the barrier pointer to reference the barrier address; and

issuing, by the scrubber, the read request in response to expiration of the time window,

wherein the COU does not receive a request to access the barrier address during the time window.

5. The method of claim 1 , wherein:

memory migration is triggered by a failure of a memory module of a plurality of interleaved memory modules; and

the plurality of interleaved memory modules has a cardinality of 2 N .

6. The method of claim 5 , wherein N is equal to 4.

7. The method of claim 5 , wherein incrementing the barrier pointer comprises adding 2 N to the barrier pointer.

8. The method of claim 5 , further comprising:

generating a data structure of addresses referring to the entries in the memory module,

wherein incrementing the barrier pointer comprises advancing the barrier pointer by one to the next address in the data structure.

9. The method of claim 5 , wherein the memory module is a DIMM.

10. The method of claim 1 , wherein the first addressing scheme is 16 way interleaved mapping, and wherein the second addressing scheme is 15 way interleaved mapping.

11. A system for migration from a first addressing scheme to a second addressing scheme, the system comprising:

a user application configured to:

issue a first request to access a first memory address; and

issue a second request to access a second memory address;

a coherence ordering unit (COU) comprising a barrier pointer referencing a barrier address and configured to:

compare the first memory address and the second memory address with the barrier address, wherein the barrier address separates migrated addresses and un-migrated addresses, and wherein the first memory address, the second memory address, and the barrier address are different;

place a first tag indicative of the first addressing scheme on the first request in response to the first memory address being on an un-migrated side of the barrier address; and

place a second tag indicative of the second addressing scheme on the second request in response to the second memory address being on a migrated side of the barrier address; and

a first memory controller (MC) connected to the COU and configured to interpret the first request according to the first addressing scheme in response to the first tag; and

a second MC connected to the COU and configured to interpret the second request according to the second addressing scheme in response to the second tag.

12. The system of claim 11 , further comprising:

a scrubber connected to the COU and configured to:

issue a read request directed towards the barrier address;

receive a value in response to the read request; and

issue a write request comprising the value and directed toward the barrier address,

wherein the COU is further configured to:

tag the write request with the second tag;

send the write request for interpretation; and

increment the barrier pointer to reference a new barrier address in response to the write request,

wherein the migrated addresses are below the new barrier address.

13. The system of claim 12 , wherein:

the scrubber comprises a timer;

the timer establishes a time window in response to the barrier pointer being incremented;

the read request is issued by the scrubber in response to the expiration of the time window; and

the COU does not receive a request to access the barrier address during the time window.

14. The system of claim 11 , further comprising:

a plurality of interleaved memory modules having a cardinality of 2 N ,

wherein memory migration is triggered by a failure of a memory module of the plurality of interleaved memory modules.

15. The system of claim 14 , wherein N is equal to 4.

16. The system of claim 14 , wherein the memory module is a DIMM.

17. The system of claim 14 , wherein the COU further comprises:

a data structure of addresses referring to the entries in the memory module,

wherein incrementing the barrier pointer comprises advancing the barrier pointer by one to the next address in the data structure.

18. The system of claim 11 , wherein the first addressing scheme is 16 way interleaved mapping, and wherein the second addressing scheme is 15 way interleaved mapping.

19. A system for migration from a first addressing scheme to a second addressing scheme, the system comprising:

a user application configured to:

issue a first request to access a first memory address; and

issue a second request to access a second memory address;

a coherence ordering unit (COU) configured to:

modify a barrier pointer to reference a barrier address;

compare the first memory address and the second memory address with the barrier address, wherein the barrier address separates migrated addresses and un-migrated addresses, and wherein the first memory address, the second memory address, and the barrier address are different;

place a first tag indicative of the first addressing scheme on the first request in response to the first memory address being on an un-migrated side of the barrier address; and

place a second tag indicative of the second addressing scheme on the second request in response to the second memory address being on a migrated side of the barrier address;

a scrubber connected to the COU and configured to:

establish a time window in response to the barrier pointer being modified;

issue a read request directed towards the barrier address in response to expiration of the time window;

receive a value in response to the read request; and

issue a write request comprising the value and directed toward the barrier address;

a first memory controller (MC) connected to the COU and configured to interpret the first request according to the first addressing scheme in response to the first tag; and

a second MC connected to the COU and configured to interpret the second request according to the second addressing scheme in response to the second tag.

20. The system of claim 19 , wherein:

the migrated addresses are above the barrier address;

modifying the barrier pointer comprises decrementing the barrier pointer; and

migration from the first addressing scheme to the second addressing scheme is triggered by failure of a memory module in a plurality of interleaved memory modules.