Well after some messing around I have eventually dug out some meaningful(?) information.
We engineered a simple test with a stored procedure that returned immediately – so the SQL is very simple (SELECT * from storedProc). This means no DB activity. On a test machine that took ~44usecs....measured from the client. This sounds good BUT when I test the UDS on the same machine the round trip latency is ~3usecs. The implication is that the cost is in the message handling – at both ends. Digging into this further I ran ltrace against the postgres server......it was seriously busy but then again my client was spinning running the same request continually. The results from ltrace shows that the server is doing what appears to be data copies – it’s doing lots of strlen/strncpy/memcpy functions. Since the stored procedure is just returning it shouts that the server costs are dominated by message handling (copying in/out of buffers). Is there any way that this cost can be reduced? As I understand it Postgres the client/server protocol is platform independent EVEN when the client and server are on the same machine. Is there an optimisation that can avoid this overhead? Another more radical solution to my admittedly very specific problem is to run the server and client in the same process.....don’t think that can be done with postgres As always any help appreciated. Cheers From: Rick Otten [mailto:rottenwindf...@gmail.com] Sent: 30 May 2017 11:18 To: Hughes, Kevin <kevin.hug...@uk.fujitsu.com> Cc: pgsql-performa. <pgsql-performance@postgresql.org> Subject: Re: [PERFORM] Client Server performance & UDS Establishing a connection with a PostgreSQL database is a very expensive process on the database server. On the other hand, establishing a connection with pgbouncer is very fast. Offloading the connection management to pgbouncer can significantly reduce the connection set up time. I've found it to help even with applications that have built-in connection pooling. If your clients are keeping persistent connections open to the database, and the latency you are experiencing is within the transaction itself, you might look at disk I/O for your WAL (write ahead logs) and take a closer look at WAL and checkpoint tuning. On Tue, May 30, 2017 at 3:34 AM, kevin.hug...@uk.fujitsu.com<mailto:kevin.hug...@uk.fujitsu.com> <kevin.hug...@uk.fujitsu.com<mailto:kevin.hug...@uk.fujitsu.com>> wrote: Hi Rick thanks for the reply. Our aim is to minimise latency hence we have a dedicated 1:1 relationship between the client and the server. If I use connection pooling surely this introduced latency – getting a server from the pool establishing the connection? Am I missing something? From: Rick Otten [mailto:rottenwindf...@gmail.com<mailto:rottenwindf...@gmail.com>] Sent: 27 May 2017 13:27 To: Hughes, Kevin <kevin.hug...@uk.fujitsu.com<mailto:kevin.hug...@uk.fujitsu.com>> Cc: pgsql-performa. <pgsql-performance@postgresql.org<mailto:pgsql-performance@postgresql.org>> Subject: Re: [PERFORM] Client Server performance & UDS You should have a layer such as pgbouncer between your pg instance and your application. It is designed to mitigate the access latency issues you describe. On May 26, 2017 10:03 AM, "kevin.hug...@uk.fujitsu.com<mailto:kevin.hug...@uk.fujitsu.com>" <kevin.hug...@uk.fujitsu.com<mailto:kevin.hug...@uk.fujitsu.com>> wrote: Hi, This is a general question around this performance area rather than a specific performance problem.....so I apologise now for a lack of a specific detail. We have an application that does many small actions on the DB – and it’s a small DB (a 50/100 Mbytes) so we would expect it to be contained in memory. Accesses need to be low latency – unfortunately there are “serial” accesses where the result of one access governs the next. Luckily the work to be done by the DB is, we believe, very simple and hence fast. Everything is running on one (large) server so we use UDS to connect the client to the server. Out observation (suspicion) is that the latency of the access, as opposed to the cost of the query, is high. Having done some investigation we believe the UDS latency may be contributing AND the cost imposed by postgres in “formatting” the messages between the client and server (transformation to network format?). We will try and get underneath this with real results/measurements but I would appreciate any comments pointers on what we are doing and how/if we can optimise this style of applications Cheers Unless otherwise stated, this email has been sent from Fujitsu Services Limited (registered in England No 96056); Fujitsu EMEA PLC (registered in England No 2216100) both with registered offices at: 22 Baker Street, London W1U 3BW; PFU (EMEA) Limited, (registered in England No 1578652) and Fujitsu Laboratories of Europe Limited (registered in England No. 4153469) both with registered offices at: Hayes Park Central, Hayes End Road, Hayes, Middlesex, UB4 8FE. This email is only for the use of its intended recipient. Its contents are subject to a duty of confidence and may be privileged. Fujitsu does not guarantee that this email has not been intercepted and amended or that it is virus-free. Unless otherwise stated, this email has been sent from Fujitsu Services Limited (registered in England No 96056); Fujitsu EMEA PLC (registered in England No 2216100) both with registered offices at: 22 Baker Street, London W1U 3BW; PFU (EMEA) Limited, (registered in England No 1578652) and Fujitsu Laboratories of Europe Limited (registered in England No. 4153469) both with registered offices at: Hayes Park Central, Hayes End Road, Hayes, Middlesex, UB4 8FE. This email is only for the use of its intended recipient. Its contents are subject to a duty of confidence and may be privileged. Fujitsu does not guarantee that this email has not been intercepted and amended or that it is virus-free. Unless otherwise stated, this email has been sent from Fujitsu Services Limited (registered in England No 96056); Fujitsu EMEA PLC (registered in England No 2216100) both with registered offices at: 22 Baker Street, London W1U 3BW; PFU (EMEA) Limited, (registered in England No 1578652) and Fujitsu Laboratories of Europe Limited (registered in England No. 4153469) both with registered offices at: Hayes Park Central, Hayes End Road, Hayes, Middlesex, UB4 8FE. This email is only for the use of its intended recipient. Its contents are subject to a duty of confidence and may be privileged. Fujitsu does not guarantee that this email has not been intercepted and amended or that it is virus-free.
