[lvs-users] persistence and destination IP hashing
David.Waddell at owmobility.com
Fri Oct 24 14:11:06 BST 2014
We are trying to use LVS as a virtual load balancer around a transparent http proxy, and are wondering if it is possible to use persistence and destination hashing together
(from our tests and as suggested in how-tos, persistence is CIP based, and therefore may not be so).
To explain our setup a little
- We have a pool of VMs running a transparent http proxy
- We have setup an LVS service on either side of this proxy with
- A 'request' path service that will schedule requests into the proxy, using src based hashing
- A 'response' path service that will schedule the http responses through the proxy, using dst based hashing.
- The http proxy maintains state around the 'session' (set of urls requested by client ip), so we wish to direct clients to the same proxy instance.
Src hashing allows to achieve this, and dst hashing on the response path ensure TCP connections get established correctly.
An issue arises when we try to add new instances of the proxy to the pool.
The hashing changes, which breaks the statefulness (user request may go to different server).
To that extent, we added persistence, which worked for requests.
However, persistence on the response path is sending , perhaps as expected, TCP packets to the 'wrong' proxy instance in a lot of cases.
This is because the persistence logic is using the web server IP address (the CIP on the response path).
An example of the problem for us :
Using 2 client ips 172.29.0.12, 172.29.0.11; real servers (http proxies) 192.168.90.58, 192.168.90.59; web server 192.168.10.17.
IP Virtual Server version 1.2.1 (size=4096) Prot LocalAddress:Port Scheduler Flags
-> RemoteAddress:Port Forward Weight ActiveConn InActConn FWM 1 sh persistent 60
-> 192.168.90.58:0 Route 1 0 0
-> 192.168.90.59:0 Route 1 0 0 FWM 3 dh persistent 60
-> 192.168.90.58:0 Route 1 0 0
-> 192.168.90.59:0 Route 1 0 0
FWM1 represents the request path; FMW 2 the response path.
IPVS connection entries
pro expire state source virtual destination pe name pe_data
A) TCP 00:58 ESTABLISHED 172.29.0.12:45659 192.168.10.17:80 192.168.90.58:80
B) IP 00:58 NONE 192.168.10.17:0 0.0.0.3:0 192.168.90.59:0
C) TCP 01:55 FIN_WAIT 172.29.0.11:50919 192.168.10.17:80 192.168.90.59:80
D) IP 00:55 NONE 172.29.0.11:0 0.0.0.1:0 192.168.90.59:0
E) TCP 00:59 SYN_RECV 192.168.10.17:80 172.29.0.12:14038 192.168.90.59:14038
F) TCP 01:55 FIN_WAIT 192.168.10.17:80 172.29.0.11:42443 192.168.90.59:42443
G) IP 00:58 NONE 172.29.0.12:0 0.0.0.1:0 192.168.90.58:0
In the example above, C and F represent a successful proxied http request, C the request from client to proxy, F the response from web server to proxy.
The request C/F was made first and establishes a persistent connections from client 172.29.0.11 ->192.168.90.59 proxy and for web server 192.168.10.17 to proxy 192.168.90.59.
We subsequently make a request A) from client 172.29.0.11, src hashing places this correctly on proxy 192.168.90.59.
The proxy then requests to web server, and the response is shown in E) - persistence directs the response to proxy 192.168.90.58, meaning a TCP connection 192.168.90.59 <-> 192.168.10.17 cannot be established.
Obviously re-engineering of the proxy instances to share state would be the ideal solution, as then persistence would not be required.
But we are wondering if there is any means to combine destination hashing and persistence successfully ? As currently persistence is enforcing scheduling based on src IP, even when dh is specified.
More information about the lvs-users