LB Delayed binding

Delayed Binding 


Syn -> LB -> ack (sent back) 

TCP packets (sIP, sP, dIP dP, seq, ack, ....)

Seq # is the # of the packet
The ack the LB has to choose a "right to left 

42 -> LB 
LB ->ACK 43 , seq -101(made up by LB)  send me more of the header
Back and forth

Syn to info server 3  seq-42  
Ack back to LB s seq 17 ack =43

Every packet going left to right : seq fiel is OK but the seq is wrong 
(it needs to have the seq # of the initial server seq # ) 

(change both direction) 

ack # ok but now the delta has be 

LB has to keep a table that maps the ID of the TCP stream and keep track of the deltas for the ack / seq to maintain the correct connections ..

Load Balancing:
- add-a-box weaknesses: 
-- singel point of failure
-- session maintenance

Failover: 
- keep a hot/warm standby : state ready machine

example Dynamic DNS - reliable, change, a name to IP binding in order of a minute TTL to 1min or 5 min .( includes waiting for client cache timeout )

100 min MTTR 
10000 min MTTF ~ .99 up time

LB - IP 1
AS 
AS
LB - IP 2

every 10000 minutes my LB fails, and in one minute it returns 
for 100min I only have one LB and no fail over.

failure every 10K minutes 1K recovery . 9999 of up time

the driver is the response of the failover..

Amazon feature: 

IP remapping - the place you want to put dynamic DNS

DNS " mysite = sIP
goes to sIP --> LB

LB fails

client cache timeout and DNS timeout before you can remap

we can us an "interface IP" to use a last hop change (sub second response) 

Can use
* round robin DNS gives me several equal bandwidth, 
* IP remapping just in front of LB  
* standby hardware for LB (customers hate this as the HW is not being use in a good state)



Load Balancing:
- add-a-box weaknesses: 
-- singel point of failure
-- session maintenance

Failover: 
- keep a hot/warm standby : state ready machine

example Dynamic DNS - reliable, change, a name to IP binding in order of a minute TTL to 1min or 5 min .( includes waiting for client cache timeout )

100 min MTTR 
10000 min MTTF ~ .99 up time

LB - IP 1
AS 
AS
LB - IP 2

every 10000 minutes my LB fails, and in one minute it returns 
for 100min I only have one LB and no fail over.

failure every 10K minutes 1K recovery . 9999 of up time

the driver is the response of the failover..

Amazon feature: 

IP remapping - the place you want to put dynamic DNS

DNS " mysite = sIP
goes to sIP --> LB

LB fails

client cache timeout and DNS timeout before you can remap

we can us an "interface IP" to use a last hop change (sub second response) 

Can use
* round robin DNS gives me several equal bandwidth, 
* IP remapping just in front of LB  
* standby hardware for LB (customers hate this as the HW is not being use in a good state)

Add-a-box weakness multiple AS and HA adn Session mgmt

stateless server (virtual homogeneity

every server holds ( or can find ) data from any session

    • full replicaition @ every application server
    • session in specialized DB (high performace server, in memory DB)
Full Replication
  • does not scale
  • spse k server
  • what is the load ?
    • every server must do msg recv and updt for evry global session update
      • total number of work increases linearly and has a ceiling
  • but with many applications , most READ but DO NOT CHANGE Session information

DB for session data

  • changed session data update DB
  • but for read only action had to read session even though it does not change anyghitng
  • then split the DB into M "shards"
    • get session data for session S in the DB and stored n M shard of DB
  • No labels