Hello, On Tue, Apr 04, 2017 at 10:54:41AM -0700, Guang Yang wrote: > I am experimenting the IO throttling with cgroups, with some good > reading from online and experiments, I found that we have two options: > 1. with CFQ IO scheduler, we can set a percentage of serving time one > group can take, and if one group does not fully take its share, other > (busy) group can steal the shares from the (ideal) group. > 2. with deadline IO scheduler, we can set an absolute IOPS and > bandwidth for one group, however, this does not consider the cost of > serving the IO (e.g. sequential vs. random) and thus it is hard to > give fair share between groups. Yeah, that's correct. > Neither one seems fully fit into our requirement - we hope that cost > model is based on the serving time, at the same time, we don't have > one group be able to steal share from the other. I see. You want time based control which isn't work-conserving. > By looking online, not sure if I missed anything, I didn't see a way > to configure cgroups IO throttling to achieve the above requirement, > besides, if it is not supported now, is there anything on the roadmap > to support it in the future? It currently isn't supported and I don't think anyone is working on implementing something like that. For SSDs, we haven't found a way to estimate the "serving time" in a scalable manner, so the outlook for the time based scheduling / control doesn't look too bright. It's a lot more cumbersome to configure but in a lot of cases limiting both bandwidth and IOPS can be good enough in limiting usage to some proportion across different workloads. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe cgroups" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
