[log] Optimize the buffer size and flush time

[Miss-you]

1. buffer=4096 is better for Writes of more than PIPE_BUF bytes may be nonatomic
2. flush=1. Since the log buffer is lowered, the flush time should also be lowered.

NOTE: Please read the Contributing.md guidelines before submitting your patch:

https://github.com/apache/incubator-apisix/blob/master/Contributing.md#how-to-add-a-new-feature-or-change-an-existing-one

Fix #1567

[membphis]

Have you done performance tests for two versions?

[Miss-you]

I will do the performance test this week

[membphis]

I will do the performance test this week

I need your result to confirm if we can merge this PR. waiting for your response.

[membphis]

@Miss-you any news?

[membphis]

LGTM

[membphis]

buffer=16384 flush=1;

I made a test, here is the result:

buffer=32768 flush=3; -> qps 15459.15
buffer=16384 flush=1; -> qps 16057.62
buffer=4096 flush=1; -> qps 14940.23

[membphis]

@Miss-you please confirm this

[Miss-you]

hi, I also made a test.

4096  Requests/sec:  16079.75
8192 Requests/sec:  16389.52
16384 Requests/sec:  16395.30
32768 Requests/sec:  16459.71

I think a log buffer size of 8192 or 16384 would be appropriate.

On the other hand, the refresh time of 3 seconds is still relatively long, and 1 or 3 seconds doesn't particularly affect QPS.

So I also agree with buffer=16384 flush=1;

[membphis]

it seems blocked. @Miss-you need your help

[Miss-you]

it seems blocked. @Miss-you need your help

I used git commit --amend to regenerate the new commitid, which should retrigger the travis