You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@trafficserver.apache.org by goran kent <go...@gmail.com> on 2011/10/17 16:51:18 UTC

ICP^2

Good day,

Have been monitoring this list with a view to switching out squid in
favour of apache traffic server on our cluster of 36 reverse-proxy
machines.  This outstanding issue of not being to use ICP amongst them
is a major concern (squid works well here - so cached content is not
duplicated).

Is there *any* plan to address this soon?

ta!

Re: ICP^2

Posted by "Henry C." <he...@cityweb.co.za>.
On Tue, October 18, 2011 16:46, Leif Hedstrom wrote:
> I don't know if the docs are clear, but clustering is superior to ICP in
> every way. That is one reason why no one has cared for ICP I think. The main
> reason for us to get ICP working is for interoperability with other ICP
> caches (e.g. Squid).
>
> Also, ICP is a flawed protocol, a better option is for us to implement HTCP,
> which Squid also supports.

Thanks for the comments, Leif.  I lockjawed onto ICP since I was most familiar
with it, assuming it was the best way to allow inter-node sharing of cached
pages.  I've been trying out ATS' clustering facility, and you're right, it's
superior.


h



Re: ICP^2

Posted by Leif Hedstrom <zw...@apache.org>.
I don't know if the docs are clear, but clustering is superior to ICP in every way. That is one reason why no one has cared for ICP I think. The main reason for us to get ICP working is for interoperability with other ICP caches (e.g. Squid).

Also, ICP is a flawed protocol, a better option is for us to implement HTCP, which Squid also supports.

cheers,

-- Leif 

On Oct 18, 2011, at 4:41 AM, "Henry C." <he...@cityweb.co.za> wrote:

> On Tue, October 18, 2011 09:24, ming.zym@gmail.com wrote:
>> 在 2011-10-18二的 08:43 +0200,Henry C.写道:
>>> On Tue, October 18, 2011 04:46, ming.zym@gmail.com wrote:
>>> Only once the down node
>>> was brought back up did things become snappy again.
>> there should be no issue when one cluster member down, it will discover the
>> change and rehash the contents to the left nodes in cluster. the timeout or
>> heatbeat checking will be <10s. after that, it will be stable as before.
> 
> Thanks, you were right - this had to do with the linux virtual server and it's
> persistence setting on the directors.
> 

Re: ICP^2

Posted by "Henry C." <he...@cityweb.co.za>.
On Tue, October 18, 2011 09:24, ming.zym@gmail.com wrote:
> 在 2011-10-18二的 08:43 +0200,Henry C.写道:
>> On Tue, October 18, 2011 04:46, ming.zym@gmail.com wrote:
>> Only once the down node
>> was brought back up did things become snappy again.
> there should be no issue when one cluster member down, it will discover the
> change and rehash the contents to the left nodes in cluster. the timeout or
> heatbeat checking will be <10s. after that, it will be stable as before.

Thanks, you were right - this had to do with the linux virtual server and it's
persistence setting on the directors.


Re: ICP^2

Posted by "ming.zym@gmail.com" <mi...@gmail.com>.
在 2011-10-18二的 08:43 +0200,Henry C.写道:
> On Tue, October 18, 2011 04:46, ming.zym@gmail.com wrote:
> > if you are using ICP inside one local cluster, there is another good solution
> > for content hashing storage, the powerfull Cluster:
> > http://trafficserver.staging.apache.org/docs/trunk/admin/cluster-howto/index.
> > en.html
> >
> > we have make some improvement in the Cluster performance in trunk, should be a
> > good solution for local net clustering.
> 
> It looks like Cluster mode has an issue if the node where an object is cached
> is down/offline.  I took one of my nodes offline (which had already cached a
> test object) as a test and the other nodes then performed quite badly (ie,
> took several seconds) - perhaps timing out? - before responding... and they
> did this consistently without improving over time.  Only once the down node
> was brought back up did things become snappy again.
there should be no issue when one cluster member down, it will discover
the change and rehash the contents to the left nodes in cluster. the
timeout or heatbeat checking will be <10s. after that, it will be stable
as before.

> 
> I also noticed they did not re-cache the object but *always* went direct to
> the origin.  Would this behaviour not be a problem if one of the nodes
> crashed?  Should an existing node then not re-cache an object to ensure
> optimal performance?
the Cluster is a expanding for storage, so if it will just help when you
do caching. be sure to make TS a working cache system before turn on
cluster.
> 
> I'm currently using 3.0.1.  You mentioned trunk - where can I find what
> improvements you've made to Cluster mode?
all cluster related project is documented in:
https://cwiki.apache.org/confluence/display/TS/Clustering
> 
> Thanks
> h
> 

-- 
zym, Zhao Yongming.
aka: yonghao @ taobao.com


Re: ICP^2

Posted by "Henry C." <he...@cityweb.co.za>.
On Tue, October 18, 2011 04:46, ming.zym@gmail.com wrote:
> if you are using ICP inside one local cluster, there is another good solution
> for content hashing storage, the powerfull Cluster:
> http://trafficserver.staging.apache.org/docs/trunk/admin/cluster-howto/index.
> en.html
>
> we have make some improvement in the Cluster performance in trunk, should be a
> good solution for local net clustering.

It looks like Cluster mode has an issue if the node where an object is cached
is down/offline.  I took one of my nodes offline (which had already cached a
test object) as a test and the other nodes then performed quite badly (ie,
took several seconds) - perhaps timing out? - before responding... and they
did this consistently without improving over time.  Only once the down node
was brought back up did things become snappy again.

I also noticed they did not re-cache the object but *always* went direct to
the origin.  Would this behaviour not be a problem if one of the nodes
crashed?  Should an existing node then not re-cache an object to ensure
optimal performance?

I'm currently using 3.0.1.  You mentioned trunk - where can I find what
improvements you've made to Cluster mode?

Thanks
h


Re: ICP^2

Posted by "Henry C." <he...@cityweb.co.za>.
On Tue, October 18, 2011 04:46, ming.zym@gmail.com wrote:
> if you are using ICP inside one local cluster, there is another good solution
> for content hashing storage, the powerfull Cluster:
> http://trafficserver.staging.apache.org/docs/trunk/admin/cluster-howto/index.
> en.html

This looks like it works as documented - thanks for saving my day!


Re: ICP^2

Posted by "ming.zym@gmail.com" <mi...@gmail.com>.
if you are using ICP inside one local cluster, there is another good
solution for content hashing storage, the powerfull Cluster:
http://trafficserver.staging.apache.org/docs/trunk/admin/cluster-howto/index.en.html

we have make some improvement in the Cluster performance in trunk,
should be a good solution for local net clustering.

I think no one is working on ICP for now, patch welcome :D

FYI

在 2011-10-17一的 16:51 +0200,goran kent写道:
> Good day,
> 
> Have been monitoring this list with a view to switching out squid in
> favour of apache traffic server on our cluster of 36 reverse-proxy
> machines.  This outstanding issue of not being to use ICP amongst them
> is a major concern (squid works well here - so cached content is not
> duplicated).
> 
> Is there *any* plan to address this soon?
> 
> ta!

-- 
zym, Zhao Yongming.
aka: yonghao @ taobao.com