DISTRIBUTED IN-ORDER LOAD SPREADING RESILIENT TO TOPOLOGY CHANGES

• US 20150095404A1
• Filed: 09/27/2013
• Published: 04/02/2015
• Est. Priority Date: 09/27/2013
• Status: Active Grant

• Current Assignee: Intel Corporation
• Original Assignee: Intel Corporation
• Inventors: Perlman, Radia
• Granted Patent: US 9,917,891 B2
• Time in Patent Office: Days
• US Class Current: 709/203
• CPC Class Codes: H04L 67/1027 Persistence of sessions during load balancing
• CPC Class Codes: H04L 67/1034 Reaction to server failures by a load balancer

Abstract

Methods, apparatus, and systems for distributing network loads in a manner that is resilient to system topology changes. Distribution functions and associated operations are implemented on multiple load splitters such that if a load splitter becomes inoperative, another or other load splitters can forward packets corresponding to flows previously handled by the inoperative load splitter without requiring flow state synchronization to be maintained across load splitters. The distribution functions are implemented in a manner that distributes packets for the same flows to the same servers through system topology changes, addressing both situations when servers fail and/or are taken off-line and when such servers or replacement servers are brought back on-line. The techniques are facilitated, in part, via use of redistributed flow lists and/or Bloom filters that are marked to track redistributed flows. A novel Bloom filter recycle scheme is also disclosed.

Claims

1. A method implemented by a load splitter configured to receive incoming packets and distribute the packets to a plurality of servers coupled to the load splitter, the method comprising:

detecting that n servers are available to receive packets;

employing an n-way distribution function to distribute packets to the n servers, wherein the n-way distribution function is configured such that packets belonging to the same flow are distributed to the same server;

detecting that a first of the n servers is unavailable to receive packets, leaving n−

1 available servers;

while there are n−

1 available servers,

in response to receiving a first packet,

classifying the first packet to associate the first packet with a flow; and

performing the n-way distribution function on the packet;

if the result of the n-way distribution function directs the first packet to be distributed to an available server, forwarding the first packet to the available server;

if the result of the n-way distribution function directs the first packet to be distributed to the first unavailable server,

employing an (n−

1)-way distribution function to determine an available server to redistribute the first packet to and forwarding the first packet to that server;

storing indicia indicating that packets associated with the flow for the first packet belong to a redistributed flow; and

when the number of available servers returns to n servers,

forwarding subsequently received packets associated with the redistributed flow to the same server the first packet was redistributed to.

2. The method of claim 1, further comprising:

in response to receiving a packet,

initializing an integer i=n;

(a) performing an i-way distribution function to identify a server to distribute or redistribute the packet to;

(b) if the result of the i-way distribution function identifies an available server, determining whether the packet is associated with a redistributed flow for the available server; and

i. if the packet is associated with a redistributed flow for the available server decrementing i by 1 and returning to (a);

otherwise,

ii. forwarding the packet to the available server;

(c) if the result of the i-way distribution function identifies an unavailable server, storing indicia indicating that packets belonging to a flow associated with the packet belong to a redistributed flow and decrementing i by 1 and returning to (a); and

(d) repeating operations (a), (b), (b)i and (c), as applicable, until the i-way distribution function identifies an available server in operation (b) that does not have a redistributed flow for which the packet is associated.

3. The method of claim 2, further comprising:

for each received packet, performing a flow classification operation to identify a flow the packet is to be associated with, the flow having a flow identifier (FlowID); and

employing the FlowID as an input to the i-way distribution function.

4. The method of claim 3, further comprising employing the FlowIDs for flows that are redistributed as the indicia indicating that packets associated with those flows are redistributed flows.

5. The method of claim 2, further comprising:

employing a redistributed flow list for a server for which flows are redistributed, wherein,

in operation (c) a redistributed flow entry is added to the redistributed flow list for the server containing indicia indicating that packets associated with the flow are redistributed to another server;

in operation (b), determining whether the packet is associated with a redistributed flow for the available server by determining a flow for which the packet is associated and performing a lookup against the redistributed flow list for the server to determine whether there is an entry matching the flow for which the packet is associated; and

in operation (b) i, updating a timestamp for the redistributed flow entry corresponding to the redistributed flow for which the packet is associated.

6. The method of claim 2, further comprising:

employing a pair of Bloom filters for storing bit vector data that is configured to facilitate confirmation to whether a flow comprises a redistributed flow for a server associated with the pair of Bloom filters;

periodically recycling a current Bloom filter by replacing it with a next Bloom filter, wherein the next Bloom filter becomes a new current Bloom filter and the bit vector data in the current Bloom filter is cleared and it becomes the new next Bloom filter;

in operation (c) bits are set in both the current and next Bloom filters associated with the unavailable server indicating that packets associated with the flow are redistributed to another server;

in operation (b), determining whether the packet is associated with a redistributed flow for the available server by determining a flow for which the packet is associated and performing a lookup against the current Bloom filter for the server to determine whether there is an entry matching the flow for which the packet is associated; and

in operation (b) i, setting bits in the next Bloom filter indicating that packets associated with the flow are redistributed to another server.

7. The method of claim 1, further comprising:

while the number of servers is less than n servers;

maintaining redistributed flow indicia containing data identifying, for each server that is currently or was unavailable, which flows are being or were redistributed away from the server to another server.

when the number of available servers returns to n servers,

in response to receiving a next packet,

classifying the next packet to associate the next packet with a flow and performing the n-way distribution function on the packet;

checking the redistributed flow indicia for the server identified by the n-way distribution function; and

,

if the flow associated with the next packet is a flow identified by the redistributed flow indicia, redirecting the packet to the server packets associated with the flow are being redistributed to.

8. The method of claim 1, further comprising maintaining at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, which flows are being or were redistributed away from the server to another server.

9. The method of claim 1, further comprising employing at least one Bloom filter for storing bit vector data that is configured to facilitate confirmation to whether a flow comprises a redistributed flow for a server associated with the at least one Bloom filter.

10. The method of claim 9, further comprising:

employing at least one pair of Bloom filters for storing bit vector data that are configured to facilitate identification of flows that have been redistributed;

for each pair of Bloom filters,

populating the bit vector data in the Bloom filters in a manner offset in time, wherein a Bloom filter that has been least recently cleared is designated a current Bloom filter and the other Bloom filter designated a next Bloom filter;

periodically promoting the next Bloom filter to become a new current Bloom filter and recycling the current Bloom filter to be employed as a new next Bloom filter, wherein recycling a current Bloom vectors clears its bit vector data; and

performing flow identification hit testing against the Bloom filter that is the current Bloom filter.

11. The method of claim 9, further comprising:

associating a pair of Bloom filters for each of a plurality of servers;

for each pair of Bloom filters,

designating a status of one of the Bloom filters a current Bloom filter and a status of the other Bloom filter a next Bloom filter, wherein the status of each Bloom filter is periodically cycled between the current Bloom filter status and the next Bloom filter status,

initially populating, for a first time period, only a first current Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server associated with the pair of Bloom filters while the server is unavailable during the first time period;

(a) for a next time period, populating both the first current Bloom filter and first next Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server during the next time period;

(b) after a next time period completes, promoting the first next Bloom filter as a new current Bloom filter and recycling the first current Bloom filter to be employed as a new next Bloom filter, wherein the bit vector data for a Bloom filter is cleared when the Bloom filter is recycled; and

repeating operations (a) and (b) on an ongoing basis.

12. The method of claim 1, further comprising:

maintaining at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, at least a portion of which flows are being or were redistributed away from the server to another server; and

employing, for each server that is currently or was unavailable, at least one Bloom filter for storing bit vector data that is configured to facilitate confirmation to whether a flow comprises a flow that is currently or was redistributed away from a server associated with the Bloom filter containing the bit vector data.

13. An apparatus, comprising:

a plurality of upstream network ports;

at least one downstream network port;

wherein the apparatus is configured to be installed in a system and to be implemented as a load splitter that distributes packets received at the at least one downstream port to a number of servers communicatively coupled to the plurality of upstream ports, wherein an initial number of servers is n, the apparatus further configured, when installed in the system and operating, to perform operations including,

in response to receiving a packet,

initializing an integer i=n;

identifying a flow the packet is to be associated with;

(a) performing an i-way distribution function to identify a server to distribute or redistribute the packet to;

(b) if the result of the i-way distribution function identifies an available server, determining whether the packet is associated with a redistributed flow for the available server; and

i. if the packet is associated with a redistributed flow for the available server decrementing i by 1 and returning to (a);

otherwise,

ii. forwarding the packet to the available server;

(c) if the result of the i-way distribution function identifies an unavailable server, storing indicia indicating that packets belonging to the flow the packet is associated with belong to a redistributed flow, decrementing i by 1, and returning to (a); and

(d) repeating operations (a), (b), (b)i and (c), as applicable, until the i-way distribution function identifies an available server in operation (b) that does not have a redistributed flow for which the packet is associated, resulting in the packet being forwarded to the available server in (b)ii.

14. The apparatus of claim 13, wherein the apparatus is further configured, when installed in the system and operating, to perform operations including:

for each received packet, performing a flow classification operation to identify a flow the packet is to be associated with, the flow having a flow identifier (FlowID);

employing the FlowID as an input to the n-way and (n−

1)-way distribution functions; and

tracking redistributed flows associated with packet that are redistributed using the FlowIDs for the redistributed flows.

15. The apparatus of claim 13, wherein the apparatus is further configured, when installed in the system and operating, to track flows that have been redistributed using at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, which flows are being or were redistributed away from the server to another server when the server is or was unavailable.

16. The apparatus of claim 15, wherein the apparatus is further configured, when installed in the system and operating, to:

add a redistributed flow entry in operation (c) to the redistributed flow list for the server containing indicia indicating that packets associated with the flow are redistributed to another server;

determine whether the packet is associated with a redistributed flow for the available server in operation (b) by determining a flow for which the packet is associated and performing a lookup against the redistributed flow list for the server to determine whether there is an entry matching the flow for which the packet is associated;

in operation (b) i, update a timestamp for the redistributed flow entry corresponding to the redistributed flow for which the packet is associated; and

remove a redistributed flow entry if the time difference between its timestamp and the current time exceeds a predetermined value.

17. The apparatus of claim 13, wherein the apparatus is further configured, when installed in the system and operating, to employ at least one Bloom filter for each server that is currently or was unavailable, the at least one Bloom filter storing bit vector data that is configured to facilitate confirmation to whether a flow has been redistributed away from a server associated with a Bloom filter.

18. The apparatus of claim 17, wherein the apparatus is further configured, when installed in the system and operating, to perform operations including:

associating a pair of Bloom filters for each of a plurality of servers operatively coupled to the upstream network ports;

for each pair of Bloom filters,

designating a status of one of the Bloom filters as a current Bloom filter and a status of the other Bloom filter a next Bloom filter, wherein the status of each Bloom filter is periodically cycled between the current Bloom filter status and the next Bloom filter status,

initially populating, for a first time period, only a first current Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server associated with the pair of Bloom filters while the server is unavailable during the first time period;

(a) for a next time period, populating both the first current Bloom filter and first next Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server during the next time period;

(b) after a next time period completes, promoting the first next Bloom filter as a new current Bloom filter and recycling the first current Bloom filter to be employed as a new next Bloom filter, wherein the bit vector data for a Bloom filter is cleared when the Bloom filter is recycled; and

repeating operations (a) and (b) on an ongoing basis.

19. The apparatus of claim 18, wherein the apparatus is further configured, when installed in the system and operating, to:

in operation (c), set bits in both the current and next Bloom filters associated with the unavailable server indicating that packets associated with the flow are redistributed to another server;

in operation (b), determine whether the packet is associated with a redistributed flow for the available server by performing a lookup against the current Bloom filter for the server to determine whether there is an entry matching the flow for which the packet is associated; and

in operation (b) i, set bits in the next Bloom filter indicating that packets associated with the flow are redistributed to another server.

20. The apparatus of claim 17, wherein the apparatus is further configured, when installed in the system and operating, to:

maintain at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, at least a portion of which flows are being or were redistributed away from the server to another server;

in response to receiving a first packet of a flow,

if the i-way distribution function for the packet identifies an unavailable server, determine whether there is any space available in the redistributed flow list for the unavailable server to add a new redistributed flow entry; and

if there is space available, add a new redistributed flow entry corresponding to the flow for the first packet to the redistributed flow list;

otherwise

add bit vector data to the current and next Bloom filters associated with the unavailable server to indicate the flow is a redistributed flow.

21. The apparatus of claim 13, wherein the apparatus further comprises:

a processor, having one or more processing elements;

at least one memory, operatively coupled to the processor;

a storage device, having instructions stored therein configured to be executed on the processor to cause the recited operations to be performed when the apparatus is installed in the system and operating.

22. The apparatus of claim 13, wherein the apparatus further comprises a management network port, and the apparatus is configured, when installed in a system employing a second apparatus having a similar configuration, to share information with the second apparatus concerning redistributed flows.

23. A tangible non-transient machine-readable medium having instructions stored thereon configure to be executed by at least one processing element on a load splitter to enable the load splitter to effect packet distribution and redistribution operations when the load splitter is installed in a system and operating, the packet distribution and redistribution operations comprising:

in response to receiving a packet,

initializing an integer i=n;

identifying a flow the packet is to be associated with;

(a) performing an i-way distribution function to identify a server to distribute or redistribute the packet to;

(b) if the result of the i-way distribution function identifies an available server, determining whether the packet is associated with a redistributed flow for the available server; and

i. if the packet is associated with a redistributed flow for the available server decrementing i by 1 and returning to (a);

otherwise,

ii. forwarding the packet to the available server;

(c) if the result of the i-way distribution function identifies an unavailable server, storing indicia indicating that packets belonging to the flow the packet is associated with belong to a redistributed flow, decrementing i by 1, and returning to (a); and

(d) repeating operations (a), (b), (b)i and (c), as applicable, until the i-way distribution function identifies an available server in operation (b) that does not have a redistributed flow for which the packet is associated, resulting in the packet being forwarded to the available server in (b)ii.

24. The tangible non-transient machine-readable medium of claim 23, wherein the instructions are configured to perform further operations, upon execution, comprising:

for each received packet, performing a flow classification operation to identify a flow the packet is to be associated with, the flow having a flow identifier (FlowID) or retrieving a FlowID associated with the packet that was generated through a previous flow classification operation;

employing the FlowID as an input to the i-way distribution function; and

tracking redistributed flows associated with packet that are redistributed using the FlowIDs for the redistributed flows.

25. The tangible non-transient machine-readable medium of claim 23, wherein the instructions are configured to perform further operations, upon execution, comprising tracking flows that have been redistributed using at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, flows that are being or were redistributed when the server is or was unavailable.

26. The tangible non-transient machine-readable medium of claim 25, wherein the instructions are configured to perform further operations, upon execution, comprising:

adding a redistributed flow entry in operation (c) to the redistributed flow list for the server containing indicia indicating that packets associated with the flow are redistributed to another server;

determining whether the packet is associated with a redistributed flow for the available server in operation (b) by determining a flow for which the packet is associated and performing a lookup against the redistributed flow list for the server to determine whether there is an entry matching the flow for which the packet is associated;

in operation (b) i, updating a timestamp for the redistributed flow entry corresponding to the redistributed flow for which the packet is associated; and

removing a redistributed flow entry if the time difference between its timestamp and the current time exceeds a predetermined value.

27. The tangible non-transient machine-readable medium of claim 23, wherein the instructions are configured to perform further operations, upon execution, comprising employing at least one Bloom filter for each server that is currently or was unavailable, the Bloom filters storing bit vector data that is configured to facilitate confirmation to whether a flow has been redistributed away from a server associated with a Bloom filter.

28. The tangible non-transient machine-readable medium of claim 27, wherein the instructions are configured to perform further operations, upon execution, comprising:

associating a pair of Bloom filters for each of a plurality of servers operatively coupled to the upstream network ports;

for each pair of Bloom filters,

designating a status of one of the Bloom filters a current Bloom filter and a status of the other Bloom filter a next Bloom filter, wherein the status of each Bloom filter is periodically cycled between the current Bloom filter status and the next Bloom filter status,

initially populating, for a first time period, only a first current Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server associated with the pair of Bloom filters while the server is unavailable during the first time period;

(a) for a next time period, populating both the first current Bloom filter and first next Bloom filter with bit vector data in response to detection of flows that are redistributed away from the server during the next time period;

(b) after a next time period completes, promoting the first next Bloom filter as a new current Bloom filter and recycling the first current Bloom filter to be employed as a new next Bloom filter, wherein the bit vector data for a Bloom filter is cleared when the Bloom filter is recycled; and

repeating operations (a) and (b) on an ongoing basis.

29. The tangible non-transient machine-readable medium of claim 27, wherein the instructions are configured to perform further operations, upon execution, comprising:

in operation (c), setting bits in both the current and next Bloom filters associated with the unavailable server indicating that packets associated with the flow are redistributed to another server;

in operation (b), determining whether the packet is associated with a redistributed flow for the available server by performing a lookup against the current Bloom filter for the server to determine whether there is an entry matching the flow for which the packet is associated; and

in operation (b) i, set bits in the next Bloom filter indicating that packets associated with the flow are redistributed to another server.

30. The tangible non-transient machine-readable medium of claim 27, wherein the instructions are configured to perform further operations, upon execution, comprising:

maintaining at least one redistributed flow list containing data identifying, for each server that is currently or was unavailable, at least a portion of which flows are being or were redistributed away from the server to another server;

in response to receiving a first packet of a flow,

if the i-way distribution function for the packet identifies an unavailable server, determining whether there is any space available in the redistributed flow list for the unavailable server to add a new redistributed flow entry; and

if there is space available, adding a new redistributed flow entry corresponding to the flow for the first packet to the redistributed flow list;

otherwise

adding bit vector data to the current and next Bloom filters associated with the unavailable server to indicate the flow is a redistributed flow.

31. A method comprising:

employing a pair of Bloom filters for storing bit vector data that are configured to facilitate identification of members of a set;

populating bit vector data in the pair of Bloom filters derived from processing element input values with a plurality of Bloom filter hash algorithms, wherein the same hash algorithms are employed by each of the Bloom filters;

populating the bit vector data in the Bloom filters in a manner offset in time, wherein a Bloom filter that has been least recently cleared is designated a current Bloom filter and the other Bloom filter designated a next Bloom filter; and

periodically promoting the next Bloom filter to become a new current Bloom filter and recycling the current Bloom filter to be employed as a new next Bloom filter, wherein recycling a current Bloom vectors clears its bit vector data.

32. The method of claim 31, further comprising:

initially populating, for a first time period, only a first current Bloom filter with bit vector data in response to processing element input values;

(a) for a next time period, populating both the first current Bloom filter and first next Bloom filter with bit vector data in response to processing element input values during the next time period;

(b) after a next time period completes, promoting the first next Bloom filter as a new current Bloom filter and recycling the first current Bloom filter to be employed as a new next Bloom filter; and

repeating operations (a) and (b) on an ongoing basis.

33. The method of claim 31, wherein an element input value comprises a Flow Identifier (FlowID).

34. The method of claim 31, further comprising determining whether an element is a member of a set by performing a Bloom filter hit test using an input value for the element against only the current Bloom filter.