Strange $upstream_response_time latency spikes with reverse proxy

[Jay Oster]

Hi again everyone!

Just posting a status update (because I hate coming across old threads with
reports of a problem I'm experiencing, and there is no answer!) What I've
found so far is starting to look like a Linux kernel bug that was fixed for
ipv6, but still remains for ipv4! Here's the relevant discussion:
https://groups.google.com/forum/?fromgroups=#!topic/linux_net/ACDB15QbHls

And thanks for making nginx awesome! :)


On Tue, Mar 19, 2013 at 3:42 PM, Jay Oster <jay at kodewerx.org> wrote:

> Hi Andrei!
>
> On Tue, Mar 19, 2013 at 2:49 AM, Andrei Belov <defan at nginx.com> wrote:
>
>> Hello Jay,
>>
>> If I understand you right, issue can be repeated in the following cases:
>>
>> 1) client and server are on different EC2 instances, public IPs are used;
>> 2) client and server are on different EC2 instances, private IPs are used;
>> 3) client and server are on a single EC2 instance, public IP is used.
>>
>> And there are no problems when:
>>
>> 1) client and server are on a single EC2 instance, either loopback or
>> private IP is used.
>>
>> Please correct me if I'm wrong.
>>
>
> If by "client" you mean nginx, and by "server" you mean our upstream HTTP
> service ... That is exactly correct. You could also throw in another
> permutation by changing where ApacheBench is run, but it doesn't change the
> occurrence of dropped packets; only increases average latency when AB and
> nginx are on separate EC2 instances.
>
>
>> What about EC2 security group - do the client and the server use the same
>> group?
>> How many rules are present in this group? Have you tried to either
>> decrease
>> a number of rules used, or create "pass any to any" simple configuration?
>>
>
> That's a great point! We have been struggling with the number of firewall
> rules as a separate matter, in fact. There may be some relation here. Thank
> you for reminding me.
>
>
>> And just to clarify the things - under "external IP address" do you mean
>> EC2
>> instance's public IP, or maybe Elastic IP?
>
>
> I'm talking about the instance public IPs. Elastic IPs are only used for
> client access to nginx. And specifically only for managing DNS. Between
> nginx and upstream servers, the public IPs are used.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.nginx.org/pipermail/nginx/attachments/20130322/03e31b10/attachment-0001.html>