Bug 2927

In the first 8 seconds, the actual bandwidth is 8 times of expected bandwidth limit.

If the bandwidth limit is 10240Kbit/s, the actual bandwidth is 10MB/s. After 8 seconds, the actual bandwidth limit will reduce to the expected limit which is 1.25MB/s.

Created attachment 3212 [details]
Fix scp bwlimit calculation.

Looking at the misc.c:bandwidth* functions I think there's two problems.

> bandwidth_limit(struct bwlimit *bw, size_t read_len)
> { [...]
>        if (!timerisset(&bw->bwstart)) {
>                monotime_tv(&bw->bwstart);
>                return;
>        }
>        bw->lamt += read_len;

The first (minor) problem: the first buffer written (16k, by default) is not accounted for at all because it returns before it gets to the code that accounts for it.

>        if (bw->lamt < bw->thresh)
>                return;

This is the main problem: bw->thresh is set to the bandwidth limit (in kbit/s), which means that for any non-trivial bandwidth limit it'll receive many writes before it even considers inserting a delay, so you get an initial burst followed by pause until the average drops back to the intended rate.

This patch moves accounting so the first block is accounted for, and sets the initial threshold such that it'll start computing the rate on the second and subsequent blocks.

Without patch:
$ ./scp -l 1000 /tmp/1MB  localhost:/dev/null
1MB                    0%    0     0.0KB/s   --:-- ETA
1MB                  100% 1024KB   1.0MB/s   00:01    
1MB                  100% 1024KB 511.8KB/s   00:02    
1MB                  100% 1024KB 341.2KB/s   00:03    
1MB                  100% 1024KB 255.9KB/s   00:04    
1MB                  100% 1024KB 204.7KB/s   00:05    
1MB                  100% 1024KB 170.6KB/s   00:06    
1MB                  100% 1024KB 146.2KB/s   00:07    
1MB                  100% 1024KB 127.9KB/s   00:08 

With patch:
$ ./scp -l 1000 /tmp/1MB  localhost:/dev/null
1MB                    0%    0     0.0KB/s   --:-- ETA
1MB                   12%  128KB 127.9KB/s   00:07 ETA
1MB                   25%  256KB 127.9KB/s   00:06 ETA
1MB                   37%  384KB 127.9KB/s   00:05 ETA
1MB                   50%  512KB 127.9KB/s   00:04 ETA
1MB                   60%  624KB 126.3KB/s   00:03 ETA
1MB                   73%  752KB 126.5KB/s   00:02 ETA
1MB                   85%  880KB 126.6KB/s   00:01 ETA
1MB                   98% 1008KB 126.8KB/s   00:00 ETA
1MB                  100% 1024KB 124.6KB/s   00:08

Thanks for the report, the patch has been applied and will be in the OpenSSH 8.0 release.

(In reply to Darren Tucker from comment #3)
> Thanks for the report, the patch has been applied and will be in the
> OpenSSH 8.0 release.

Hi, I also submitted my patch to resolve this problem last month. However, I may misunderstand the algorithm for limiting the bandwidth, so I think my patch is wrong after reading you explanation. :)

This is my patch :)

https://lists.mindrot.org/pipermail/openssh-unix-dev/2018-November/037346.html

Move resolved bugs -> CLOSED after 8.0 release