If the main reason for contemplating this design is the size of the task payload, have you considered a content addressable storage design? For example: why can the task not be launched with URIs that allow the agent to download the payload before launching the task? This fans out the network load evenly among agents.
— *Joris Van Remoortere* Mesosphere On Wed, Feb 17, 2016 at 9:35 PM, Suteng <[email protected]> wrote: > Alex, > We don't have test the performance of mesos. But we have develop a > framework in house, which is like a simplified mesos, use to schedule a > large number fine grain computation tasks. We find that master will be a > bottleneck. One reason is our task contain several KB data, and task number > is quite huge. > If we use mesos to replace it, maybe master still be a bottleneck. > > Master still do the resource bookkeeping, we can decompose launch task to > two steps, firstly scheduler tell master which offer he wants, then master > tell scheduler the address of agent. Secondly, scheduler can directly > launch task to the agent, and also can directly send message to agent. > Maybe I can do some test about the mesos master launch task throughput, > with different number task data. > > > -----邮件原件----- > 发件人: C Rukletsov [mailto:[email protected]] > 发送时间: 2016年2月17日 18:04 > 收件人: dev > 主题: Re: Question about "Framework directly access Meso agent" > > Suteng— > > such optimization makes sense in certain cases (e.g. sending a framework > message), but it can be rather tricky in general, because the master has to > maintain bookkeeping. Moreover, with the upcoming HTTP API it becomes > harder for a framework to determine where to send messages to reach a > specific agent. > > Have you done any performance tests and seen master becoming a bottleneck? > > On Wed, Feb 17, 2016 at 5:14 AM, Suteng <[email protected]> wrote: > > > Hi, > > > > > > > > Currently, Mesos framework’s task related operations lauchTask, > > updateStatus and executorSendMessage etc., and resource related > > operations resourceOffer etc., all operations are pass through Mesos > Master. > > > > When the cluster and task number become huge, or with optimistic > > resource offer, multi-framework concurrently launchTask, maybe Mesos > > Master will be a bottleneck. > > > > Is possible for framework scheduler directly access Mesos agent, > > launchTask, updateStatus and SendMessage2Executore to Mesos Agent > > directly, bypass the Master? > > > > Will invoke big conflict with current mechanism? > > > > > > > > Looking forward to your comments and opinions. > > > > > > > > Best Regards, > > > > Teng > > > > > > > > > > > > > > > > Su Teng 00241668 > > > > > > > > Distributed and Parallel Software Lab > > > > Huawei Technologies Co., Ltd. > > > > Email:[email protected] > > > > > > > > > > >
