Often see cloud platform, cloud storage that merge the same disk IO, on the same disk IO in the order of adjustment, optimize the disk IO,
I want to ask is, how can you allow direct control of the disk IO???????
Can anybody give an example?
And control, such as memory, this is not operating systems control? You can directly control?

CodePudding user response:

Here I mainly talk about pangu we distributed storage System, the service dispatch System some specific details of architecture, the two modules of pangu architecture has many similarities with Hadoop, today I know Hadoop System can automatically switch, switch within a minute, in order to ensure that all data is not lost, we check the whole data in all links, we reached 1 billion scale, still trying to keep up, inside the pangu distributed file systems a feature we biggest store of distributed I/O lines and the background management to separate, why do you speak? Is known on the common platform we have random I/O, sequential I/O, transactional I/O, in fact, different application of I/O features, such as who will responsible for leading to copy, the client don't know this logic, the client is response quickly, can quickly adapt to the system to adjust the holes, and replication strategy, chain or a new type of replication, the chain is the use of the bandwidth is wide, the new copy response speed is faster, the last is persistence of some of the requirements, the application of transaction class you want to write a record fall plate, they don't care about so fast for large data processing level of persistence, how much they need to delay within data persistence, on the client side, we provide different types of documents to adapt to different needs, their common use public storage management platform, really fell dish how to deal with the I/O, how do we deal with priority, do redundant recovery, we put these on which machines make sure he get better data reliability,

How to deal with this on a single node I/O classification event driven I/O a management style, all the I/O requests to a queue, there is a thread pool, each corresponding to a hard disk, we ensure that a single line of concurrency, received the request queue waiting queue inside, to do so because we want to ensure priority, we ensure the QS, low application can get some bandwidth, the same priority we can prioritize reduce I/O requests the path overhead, the benefits of this architecture is the control of the physical disk complications, and can enjoy many concurrent client request,

CodePudding user response:

http://www.csdn.net/article/2013-12-05/2817724-bdtc2013-aliyun
Ali flying platform total architects Tang Hong: flying open platform

For example this speech,

CodePudding user response:

Good difficult problem, pass by to help

CodePudding user response:

Do these people operating system configuration, write your own hardware driver

CodePudding user response:

Top,,, look forward to someone to answer