On 2/6/21 6:06 AM, Atul Kumar wrote:
Hi Gavan,
Thanks for providing the details, I need more clarification on this as how should I analyze that what should be ideal no. of connections should we set to avoid IO overhead based on the available hardware resources.How to do this calculation ?
Run "iotop -o -u postgres", and then compare that with the total bandwidth available to the system. If it's (even almost) saturated, then everything else will be starved.
The "--jobs=" value should AT MOST be some *small* multiple of the number of CPUs (like 1x, 1.5x or maybe 2x if the core count is low, and nothing else is running on the system.
Note: even during 300 threads, my RAM utilisation is totally normal.
RegardsAtul
On Saturday, February 6, 2021, Gavan Schneider <list.pg.gavan@xxxxxxxxxxx> wrote:
On 6 Feb 2021, at 3:37, Ron wrote:
On 2/5/21 10:22 AM, Rob Sargent wrote:A reasonable conjecture… though there is the consideration that 300 CPU intensive tasks spread across a given number of CPUs is going to waste some resources with context switching., i.e., need more CPUs :)
On 2/5/21 9:11 AM, Ron wrote:
Obviously... don't use 300 threads.No, no Ron. Clearly the answer is more CPUs
I hope you're being sarcastic.
Basically if there is plenty of wait time for I/O completion then CPU task switching can get more total work done. So far so obvious. In this thread I can see where it is disappointing to have a system considered capable of 700 connections getting saturated by a “mere” 300 threads. But this is only a “problem” if connections are equated to threads. PG max connection count is about external users having access to resources needed to get a task done. Like all resource allocations this relies on estimated average usage, i.e., each connection only asks for a lot of CPU in brief bursts and then the result is transmitted with a time lag before the connection makes another CPU demand. The system designer should use estimations about usage and load to budget and configure the system, and, monitor it all against actual performance in the real world. Of course estimates are a standing request for outliers and the system will show stress under an unexpected load.
So far I have not seen an analysis of where the bottle neck has occurred: CPU RAM HD and/or the data bus connecting these. Some of these hardware resources maxed out to the extent the system would not immediately pick up an additional work unit. As I see it OP started 300 CPU intensive tasks on hardware intended for 700 connections. If the connection count was designed with say 50% CPU intensive time per connection you would expect this hardware to be fully saturated with 300 CPU intensive tasks. More than that, doing the task with 300 threads would probably take longer than (say) 200 threads as the increased CPU context swapping time is just wasted effort.
OP now has a choice: decrease threads or (seriously) upgrade the hardware. We in the gallery would love to see a plot of total time to completion as a function of threads invoked (50-300 increments of 50) assuming the starting conditions are the same :)
Gavan Schneider
——
Gavan Schneider, Sodwalls, NSW, Australia
Explanations exist; they have existed for all time; there is always a well-known solution to every human problem — neat, plausible, and wrong.
— H. L. Mencken, 1920
--
Angular momentum makes the world go 'round.
Angular momentum makes the world go 'round.
