Agree 100% on what you're saying here. And if this was a permanent load, this would be the route I'm taking. But I'm talking about supporting this level of load for a very short period of time (we're merging a couple datacenters and having to reinstall every server all at once, minimizing the customer impact and downtime, so speed is king).
So while setting up load balancers, scaling out the hosts, etc. are all a great idea for a long-term solution, it's too much work for a single weekend. That's why I'm going for a scale-up (we have the server and 10Gbit network already, not a load balancer). The main thing I'm asking is if anybody has had any experience scaling up a cfservd to this much work, and whether or not it fell over. Paul Krizak 7171 Southwest Pkwy MS B200.3A MTS Systems Engineer Austin, TX 78735 Advanced Micro Devices Desk: (512) 602-8775 Linux/Unix Systems Engineering Cell: (512) 791-0686 Global IT Infrastructure Fax: (512) 602-0468 On 04/30/10 11:24, michoski wrote: > On 4/30/10 8:30 AM, "Tim Cutts"<t...@sanger.ac.uk> wrote: >> On 30 Apr 2010, at 4:06 pm, Paul Krizak wrote: >> >>> Has anybody out there ever tried scaling up a cfengine server (v2.1 or >>> v2.2) on a really big, fast server? I'm thinking on the order of 4 >>> sockets, 24 cores, and a 10Gbit NIC. >>> >>> This is to support a particularly massive (and temporary) flood of >>> cfagent requests to synchronize their local policy. It's going to be a >>> lot easier to scale the server up in this case rather than adjust the >>> policy to distribute requests to multiple cfservd's. >> >> How many clients are you talking about? And how much policy? I have 2300 >> clients updating policy once an hour from a small 1GigE-connected, dual >> socket >> server (four cores total) which also runs Splunk and nagios, so is quite busy >> with other things, and it copes just fine, with a load average of 0.38. >> Total >> size of all policy files on my setup is 2.9 MB. cfengine version is 2.2.8. >> The SplayTime is also one hour, so the cfengine load on the server is more or >> less steady. > > If you've got a policy or a script that builds your cfengine servers (and > you should), it's not hard to build more cfservds (well, technically, > cfservd is usually running everywhere...and all our hosts are clients and > servers...but you know what I mean). > > So... Why not stick a few of them behind a load balanced VIP? DSR would be > best in this case, since it would off-load return traffic and scale the > NETWORK INTERFACE, RAM, etc (not just add cores). > > This is generally how you scale throughput for any other server farm > (youtube.com doesn't run on one massive server). I've said it before -- > load balancers can be cheap and free. I've supported some very popular > ecommerce sites using nothing but commodity hardware and OSS (I grumbled a > lot, but it worked). > > _______________________________________________ Help-cfengine mailing list Help-cfengine@cfengine.org https://cfengine.org/mailman/listinfo/help-cfengine
