Poor network performance (host72)
I'm a new Linode user, and so far I'm pretty happy with my new toy. Usuability is really good, excepted for one point: I rarely achieve data transfers superior to 200KB/sec. This is especially annoying when doing an apt-get dist-upgrade, and I wonder if this is a kind of voluntary bandwidth limitation, or just my host being slow. I have a linode 128 on host72.
Any performance report would be appreciated.
Thanks,
Alex.
6 Replies
Fremont (Host19): 183kbyte/sec average, 200kb/sec max.
Atlanta (Host77): 250kbyte/sec average, 254kb/sec max.
Gotta say that this looked good in 2004, but this seems to be quite miserable performance for 2007 – even more so since I acquired a 20Mbit/sec download rate with a new ADSL account a couple of months ago … .
Cliff
@c1i77:
Unfortunately ~200kb/sec seems par for the course. This is what I got with a 30MB test download from a couple of Linode128s last week:
Fremont (Host19): 183kbyte/sec average, 200kb/sec max. Atlanta (Host77): 250kbyte/sec average, 254kb/sec max.
Thanks for your answer and for testing this out. I wanted to ensure this was "normal".
> Gotta say that this looked good in 2004, but this seems to be quite miserable performance for 2007 – even more so since I acquired a 20Mbit/sec download rate with a new ADSL account a couple of months ago … .
I have to agree here. I run a couple of sites that don't receive much traffic (200 visits/day max), and I switched to Linode from my DSL connection because the latter had 1Mb upload. I'm a little bit disappointed to see that I haven't gained much there. And it seems that my Linode would have a hard time handling traffic peaks.
Anyone who knows why the network performance is so slow?
@Gnurou:
Anyone who knows why the network performance is so slow?
My theory: these network performance measurements include the effects of Linode disk i/o contention - a known bottleneck.
Justification: host to host transfers (where the host Linux is doing the disk accesses rather than the Linode) are much faster (Cliff's figures for Fremont to Atlanta were ~660 to ~830 kB/s).
Remember, your internet traffic goes through a lot more networks than just your cable modem and Linode's.
The limiter would only have an effect if it was limiting, or if the host was busy. Easy to look at /proc/io_status and see the former, and we've had little problems of late with people swap thrashing, so I wouldn't lay a blanket claim that Linode transfers are slower because of the limiter.
Our last 30 hosts that have gone online are all hardware raid, and I've yet to see anyone complain of someone swapthrashing on those hosts, even when there were swap thrashers.
I'll set up some test files for download at various locations…
-Chris
Also - tracerouting the hosts in Atlanta (host77.linode.com and up) frequently shows a noticeable ping increase between the final two hops, for example:
5 gnax.ge2-3.br01.atl01.pccwbtn.net (63.216.31.158) 96.269 ms 96.222 ms 96.266 ms
6 atl-core-e-gi4-4.gnax.net (209.51.131.30) 96.033 ms 96.180 ms 96.038 ms
7 atl-core-a-tgi2-1.gnax.net (209.51.149.105) 96.328 ms 96.717 ms 96.522 ms
8 l3-atl-18.gnax.net (209.51.131.74) 99.183 ms 99.251 ms 99.276 ms
9 63.247.71.196 (63.247.71.196) 131.531 ms 143.354 ms 142.344 ms
In this case the host is responding ~40ms slower than the hop before it, but it varies between 0ms and 100ms. Has anyone experienced this with their Atlanta Linodes?