Vlan

[alexr186]

Has then been any issues with running a grid's back-end on a gigabyte vlan?

Thanks

Alex

[PeterNic]

Alex,

What do you mean by 'gigabyte vlan'?

If you are asking whether having a VLAN on the back-end will work, the short answer is 'yes, for now'. The rest of the requirements (single switch, etc.) remain in effect.

To give you a more specific answer, please let me know what the purpose of vlanning the back-end switch is.

(This thread is probably more appropriate for the maintainers' forum -- let me know if this is so)

Regards,
-- Peter

[alexr186]

I guess I miss typed (spell check in firefox), I meant gigabit, as in the speed of the switch. It would be to create a private network on a back-end switch, that is also being used for other back-end traffic like vpns and backups. Yes it would be only one switch, all the machines would be next to each other. Can you elaborate on your answer some more.

Feel free to move the thread to were you think it would be best placed.

[PeterNic]

I guess I miss typed (spell check in firefox)

OK, I just wanted to make sure it is not a brand name. Spell checkers are dangerous - they have probably caused more trouble than some malware :) (and, yet, they are damn convenient and irreplaceable by now)


It would be to create a private network on a back-end switch, that is also being used for other back-end traffic like vpns and backups. Yes it would be only one switch, all the machines would be next to each other. Can you elaborate on your answer some more.


The current versions of AppLogic, 1.2.14 and 2.0.2, will work with VLANned back-end switch. In fact, the current scalability test grid (100+ servers) uses a VLAN on a large switch.

Still, here are some of the reasons not to do it and why we don't support this configuration in production use:

the grid backbone switch is not really used as a regular network switch. Functionally, it is a cluster interconnect and storage network -- as Myrinet, FibreChannel or Infiniband would be, except it is commodity (i.e., inexpensive, easy to find and (nearly) everyone knows to use it). Any advanced configurations that don't take this into account -- VLAN, QOS, Layer 3 switching, filters, and the myriad of useful features switch vendors add -- can wreak havoc and introduce hard-to-find failures and bottlenecks. Keeping the raw switching capacity and low latency is what the grid needs from the switch
normally, network switches can be reset -- applications easily re-establish most Internet-type connections and tolerate drops (as most of such network connections go through a complex set of networking gear). Grids use the switch for heartbeat between the controller and all of the servers; even a short-time traffic stop -- which normally won't happen on a raw dedicated switch -- can trigger various failover mechanisms. Sharing the switch between the grid and other functions makes it so much more likely for this to occur.
configuring VLANs adds further complexity and makes the configuration more prone to human error
for those who run multiple grids, AppLogic supports up to 30 grids on the same backbone switch, transparently, securely and without having to configure VLANs
we are planning to add software-based VLANs and other features that may interfere with non-raw configurations (e.g., if we try to use the NIC-supported VLAN tagging, this will conflict with the VLAN configuration on the switch).
last but not least, switches are inexpensive; you can afford to buy a dedicated switch for the backbone; both Dell and D-link offer 48-port non-blocking GigE switches for under $700 (<$15/port)


I don't know if this helps... if not, let's continue the topic. Anyone who already maintains grids is welcome to comment, too.


Feel free to move the thread to were you think it would be best placed.

I moved this thread to the grid maintenance section -- most users who use hosted grids (as private datacenters, VPS or other) don't need to worry about this. Only those who host grids on their own hardware do. Grid's end-users -- those who build and manage applications -- shouldn't care about the hardware configuration at all; this removal of concern is one of the important value propositions of grid and utility computing.

Regards,
-- Peter