Database servers tend to be fairly lightly loaded, as compared with
file servers, so the most obvious thing you want to plan for is a minimum
of down time.  The advice I got when setting up our first cell here at ucsc
was that the number wasn't so very important, as long as it was an odd one.
We selected 3 as our number so that if any one of the database servers was
down or off the net, we'd still have quarum and could continue operating
through the outage.  This has proved useful as well as it lets us perform
maintenance on the database servers during regular hours without affecting
the operation of the cell.
-Brian
On Mar 11,  5:39pm, Paul Blackburn wrote:
} Subject: Optimal number of AFS database servers?
} Hello,
} 
} If there is one, what is the optimal number of AFS database servers
} to have in a growing cell?
} 
} What are the factors which determine this?
} (I am assuming pure database servers. Eg they don't also fileserve.)
} 
} 
}    +-number of database servers (N) -------------------------------------[1]
}    |
}    |   +-quorum (eg need Q out of N servers to elect sync site)----------[2]
}    |   |
}    |   |   +- 1 server failed, chance of client connecting to server OK--[3]
}    |   |   |
}    |   |   |                 + "diminishing return"----------------------[4]
}    |   |   |                 |
} 
}    1   1   0/1   ->  0%
}    2       
}    3   2   2/3   -> 66%     66
}    4    
}    5   3   4/5   -> 80%     14
}    6    
}    7   4   6/7   -> 85%      5
}    8    
}    9   5   8/9   -> 88%      3
}   10    
}   11   6   10/11 -> 90%      2
}   12
}   13   7   12/13 -> 92%      2
}   14
}   15   8   14/15 -> 93%      1
} 
} Other factors:
} 
} a) It seems that it is best to have an odd number of db servers so
}    that in the event of the sync site server failing the remaining
}    db servers can elect a new sync site.
} 
} b) Also, while it might seem to be good to have 11 db servers
}    (so that clients have a 90% chance of connecting to a working
}    server) what is the increased probability of a machine failing
}    having 11 instead of (say) 5?
} 
} c) The more db servers you have, the better the distribution
}    of processing load.
} 
} d) There is a "diminishing return" [4] as the number of servers grows
}    Eg, growing from 1 to 3 servers yields an increase of 66% [3]
}    while growing from 9 to 11 servers yields an increase of 2% [3]
} 
} e) By having "pure" db servers, better performance and reliability
}    is obtained (eg on db servers, switch off all surplus daemons
}    and servces).
} 
} Ignoring the obvious question of cost of machines, does anyone
} have a way to compute the optimal number?
} 
} It seems to me that 3 is a good number but 5 might be best.
} -- 
} paul                             http://acm.org/~mpb
} 
}     "Young man, in mathematics you don't understand things,
}      you just get used to them." -- John van Nuemann
} 
}     "Without a quorum, we have a quandary." --unknown
} 
>-- End of excerpt from Paul Blackburn


Reply via email to