Re: [jgroups-users] Flatting the bandwidth/packet curve

On 03.04.20 22:32, Questions/problems related to using JGroups wrote:
> I'm doing some experiments with jgroups over crummy network conditions 
> including high latency and about 10% loss, just to see how it performs. 
> I'm running pretty much the sample udp.xml with increased timeouts.

How do you simulate the loss? iptables or DISCARD? Is it always 10%, or 
do you randomly drop messages in the range [0 .. 10%]?

Are these multicast (msg.dest=null) or unicast (msg.dest != null) messages?

> When transferring a larger amount of content and plotting the packets 

How large?

> per second over time, I'm noticing that I'm getting spikes of around 
> 2500-3000 packets per second every 10 seconds and then basically nothing 
> in the in between the peaks. This is far from optimal and is running 
> about 15x slower than it should be.

What's "should be"? Bandwidth/message-size ~= number of messages to be 
expected?

> a) i'm not sure what's up with the emission spikes but i'm guessing that 
> jgroups is throttling itself for some reason. if you can provide some 
> recommendations or things check or if needed modifications

* What bundler do you use (TransferQueue is the default)?
* Did you set {mcast,ucast} recv/send_buf_size in UDP?
* What size are your messages?
* What are your UFC/MFC credits?

> b) Is there any configuration tweaks i can make to tell jgroups "expect 
> about this much bandwidth" or "don't exceed X packets per second" or 
> something like that? This could help flatten the curve packets per 
> second curve out a bit.

There's an experimental protocol RATE_LIMITER that I once wrote, but I 
never got around to polishing it up, let alone maintain it. This would 
be used _instead of_ UFC/MFC

> any help is appreciated
> 
> 
> 
> _______________________________________________
> javagroups-users mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/javagroups-users
> 

-- 
Bela Ban | http://www.jgroups.org