Sags network from small loads?

Doing stress testing of a server using Apache Benchmark Tool and Apache jMeter. At 30 and above competitive query starts periodically (about every 100-200 request) sags network for 3-4 seconds.

With all of this, the load on the server, no no, LA is zero, the memory is enough.

ie 200 requests go with a 50ms response, and then ~10 go with a 3000 MS response, and then the situation repeats.


And it does not depend on the content for the tests even on static.


Config server:

Leaseweb

Intel Xeon X3440/8GB/2x1TB (QC22)

QC X3440

8GB DDR3

1 x 100Mbps Full-Duplex


What could be the problem and how can otmonitoreny weaknesses?
October 8th 19 at 03:21
4 answers
October 8th 19 at 03:23
C what makes you think that slowing down the network? Try it in parallel with the stress test (full) to ping the server. If under load pings are stable, then the network is all right. And Apache may have 100 reasons why it's slow.

Posomtrite the output from top load to begin with, how much time it sits in the core of the processes above, how to use the swap.
Also, regardless of the type of network card, the host may limit the channel width. - antonette.Gislas commented on October 8th 19 at 03:26
With regard to the limitation of the width, for example, I'm using ab tested the server, and the channel was only 10 Mbps (+ restrictions from my ISP), the page size of 40 KB, it's only 30 pages per second, I also wondered why free CPU time there is, and the server does not use it. - antonette.Gislas commented on October 8th 19 at 03:29
October 8th 19 at 03:25
check ping-ω (L3) and TCP level (L4)

if everything is OK, then dig in L7 (apache)
October 8th 19 at 03:27
put some kind of monitoring system server, such as munin, cacti, zabbix, and so on the charts you can see what resources are missing.
you have not written anything about on the server. There you have the apache spinning? Periodic failures can be related to MaxRequestsPerChild — like...
The problem is not ON.
Tested initially on Node.JS the load is zero, Node.JS on each core, the load is perfectly distributed.
Then for testing purposes, put Apache and check how will be given to statics, the same failures with a network that was Node.JS it is logical to assume that the problem is somewhere at the network level, how does it track?

With less than 30 competitive queries there are no failures. - antonette.Gislas commented on October 8th 19 at 03:30
Connect the machine with which the test directly to the server cross cable, test. Further, we dance from the result. If again the same thing, look for a problem in the network card (the network card drivers), if all is in order, to look for a problem on network equipment by replacing intermediates one at a time and testing. - antonette.Gislas commented on October 8th 19 at 03:33
then you can offer, if it is possible to stop the server, load it with what ever Resсue system is minimal under it to start nginx and test on the static, if sagging will — to blame the hardware... - Bobby.Boy commented on October 8th 19 at 03:36
Connect direct cross cable is not possible, we can rent a server at Leaseweb.
I think I will write Leaseweb to diagnose the problem. - simeon_Fay commented on October 8th 19 at 03:39
October 8th 19 at 03:29
>periodically (about every 100-200 request) sags network for 3-4 seconds.
what you describe sounds like the rebirth of the worker.

"MaxRequestsPerChild controls how frequently the server recycles processes by killing old ones and launching new ones."
We were doing server stress testing without Apache at all. We later put Apache in order to test whether there will be a problem. So I don't see how there can be employed MaxRequestsPerChild.

It is planned that the server will work Node.JS - antonette.Gislas commented on October 8th 19 at 03:32

Find more questions by tags JMeterA/B testingComputer networksIntel