LI Daobing wrote:
4. a new AFR model: Currently, if the AFR have 3 child xlators, and each xlator connect to a distinct machine. Then the write speed of this AFR is only 33% of the capacity of the network. Consider a different model, the AFR send data to machine1, machine1 send the data to machine2 immediately and then write the data to disk. Machine2 also send data to machine3 immediately and then write data to disk. Under this model, we can increase the write speed to 3 times of the previous model (if your switch is good enough and your network support full duplex).
This can be achieved through multiple AFR definitions, different for each server, that chain the data. I've discussed it once or twice on the list before, but never implemented (time doesn't permit me to follow up on lots of ideas). No need for a new translator, which would be complicated (see below).
In more detail, we need two new kinds of xlators. The first one is the combination of the AFR and client-protocol (called *safr*). The second one is similar with the `server-protocol'(called *sserver*). The machine1, machine2, machine3 is set in the option of safr. And safr maintain an active-machine list. When safr receive a writev command(or other commands), it pick a machine from the active-machine list(for exmaple, machine1). then send the data and a list "[machine2, machine3]" to machine1. machine1 forward data and list "[machine3]" to machine2 immediately, machine2 also forward data and an empty list to machine3. Machine1, machine2, machine3 also write the data to the disk when sending data. If any machine is down, the afr just remove it from the active-machine list. And add it when the machine is up again. This model is a little far from the current framework, but I think it's a good idea to write at 100 MB/s instead of 30+ MB/s in a gigabyte network.
This all assumes a lot about the target network, which is a bad thing when trying to be as flexible as GlusterFS. For example, machine1 and machine2 may not have an optimized path to each other. They may be in different subnets, off different switches, or in different geographical locations, etc.
The only thing your translators provide that isn't already available through chained translators is automatic reconfiguration of the chain members when a server drops out, which is a good feature, but I would rather just add cheap redundant hardware to boost speed, such as extra gigabit NICs and switches to allow dedicated paths between select systems. Also, maybe the new switch translator can be added to what's already available to achieve what you want, I'm still fuzzy on exactly what it can be used for.
This model is similar with the model in google file system, you can check the figure 2 in a paper of google file system[1]. I put a copy of this figure at [2]. [1] http://labs.google.com/papers/gfs-sosp2003.pdf [2] http://picasaweb.google.com/lidaobing/Public/photo#5150803669289886370
Google has a advantage of designing for a specific usage, which makes some design choices better for them than for other projects.
PS, should I copy this feature request to wiki? Or it's ok to only put it here?
OK, now that I've done my best to tear down your proposal and say why it's not needed, here's where I put my disclaimer:
1) I'm not a dev, and I haven't really looked into the code, so I don't know how easy or hard your proposal is to actually implement. 2) I'm just one person, and even though *I* may think it's not needed, others may differ on this point.
-- -Kevan Benson -A-1 Networks
