Обсуждение: Advice on running two database clusters on one server machine
Hi all, We are implementing a hospital information system and a human resources/payroll processing system on two identical dedicated servers with two Xeon Quad Core processors and 32 GB RAM each, both servers being attached via FC to a SAN, and both applications running on PostgreSQL 8.3 / CentOS 51. We are wondering about the advisability to distribute the databases between the two server machines, both machines acting as active production systems for one application each, and as warm standby servers for the other, using WAL shipping to a second database cluster running on another port on each of the two server machines. What would be the performance cost of doing so, rather than running all databases on one database cluster on one machine, and using the second machine as a warm standby server for all databases of the two applications? What other considerations should we take into account? We have no prior experience with PostgeSQL administration, having run our previous systems on Windows Servers and MS SQL Server. Thanks to all for your input! -- Andreas Philipp Clínica Universitaria Teletón Chía, Colombia
On Sun, Jun 15, 2008 at 11:11 AM, Andreas Philipp <andreas.philipp@clinicauniversitariateleton.edu.co> wrote: > Hi all, > > We are implementing a hospital information system and a human > resources/payroll processing system on two identical dedicated servers with > two Xeon Quad Core processors and 32 GB RAM each, both servers being attached > via FC to a SAN, and both applications running on PostgreSQL 8.3 / CentOS 51. > > We are wondering about the advisability to distribute the databases between > the two server machines, both machines acting as active production systems > for one application each, and as warm standby servers for the other, using > WAL shipping to a second database cluster running on another port on each of > the two server machines. > > What would be the performance cost of doing so, rather than running all > databases on one database cluster on one machine, and using the second > machine as a warm standby server for all databases of the two applications? Well, when both machines are working, your performance would be better on two machines than on one. But after a failover, the warm standby will be running two instances of postgresql, and that's sub-optimal. > What other considerations should we take into account? We have no prior > experience with PostgeSQL administration, having run our previous systems on > Windows Servers and MS SQL Server. Also, worry about the possibility of switching the wrong databases when things are going wrong, etc with this type of setup. However, the biggest concern is your FC-SAN setup. Some are very fast. Some are quite pokey, and once you buy them, the manufacturer will just laugh and point their fingers at linux, postgresql, or the phase of the moon as the cause of your performance problem and do nothing more. If you haven't tested your FC-SAN setup with both servers running comlex benchmarks (multiple bonnie++ instances, lots of dds, etc...) and proven that it's fast, don't expect it to be.
>>> "Scott Marlowe" <scott.marlowe@gmail.com> wrote: > Andreas Philipp <andreas.philipp@clinicauniversitariateleton.edu.co> wrote: >> We are wondering about the advisability to distribute the databases between >> the two server machines, both machines acting as active production systems >> for one application each, and as warm standby servers for the other, using >> WAL shipping to a second database cluster running on another port on each of >> the two server machines. >> >> What would be the performance cost of doing so, rather than running all >> databases on one database cluster on one machine, and using the second >> machine as a warm standby server for all databases of the two applications? > > Well, when both machines are working, your performance would be better > on two machines than on one. But after a failover, the warm standby > will be running two instances of postgresql, and that's sub-optimal. We have many machines here running multiple PostgreSQL clusters, including one running two clusters, each with hundreds of GB of data, which run tens of millions of queries per day without performance problems. (It's also running several Java processes through which all this traffic passes.) It is a matter of having adequate hardware, properly tuned databases, efficient software design, and well written queries. -Kevin
On Sun, Jun 15, 2008 at 12:11 PM, Andreas Philipp <andreas.philipp@clinicauniversitariateleton.edu.co> wrote: > Hi all, > > We are implementing a hospital information system and a human > resources/payroll processing system on two identical dedicated servers with > two Xeon Quad Core processors and 32 GB RAM each, both servers being attached > via FC to a SAN, and both applications running on PostgreSQL 8.3 / CentOS 51. > > We are wondering about the advisability to distribute the databases between > the two server machines, both machines acting as active production systems > for one application each, and as warm standby servers for the other, using > WAL shipping to a second database cluster running on another port on each of > the two server machines. > > What would be the performance cost of doing so, rather than running all > databases on one database cluster on one machine, and using the second > machine as a warm standby server for all databases of the two applications? > > What other considerations should we take into account? We have no prior > experience with PostgeSQL administration, having run our previous systems on > Windows Servers and MS SQL Server. > > Thanks to all for your input! I've experimented a bit with this. Probably the biggest thing to keep in mind is that different clusters don't play nice with resources, especially shared memory. You're ostensibly cutting your available memory in half by running two clusters on one machine. Other things to keep in mind...they can't use the same user and group data (so roles and passwords may be different and it's a bit of work to keep them sync'd up if you would want that). You also have to maintain two different sets of configs, have data located at two different places, and listen to data on two different ports. It's about double the basic administration. I would recommend one cluster per machine for production machines, if it works well for you, that is. Peter