The computer crosses the data processing with kernel sensitive to time allocation newly, can answer_News

The data center of nowadays can waste a large number of energy, and can waste the request that much energy will come to answer an user quickly as far as possible, and defer has a few small seconds only. Masschusetts Institute of Technology (a kind of when MIT) researcher develops new system, through crossing central processor better (CPU) kernel allocates the data processing with sensitive time, ensure hardware is efficient move, improved the efficiency that high speed operates. Data center moves as distributed network, a large number of Web and shift application program implement on individual server. When the user sends a request to applied process, the data extract of memory will from hundreds or the extraction on thousands of server. In send before answering, applied process must await the rudest service to process data, time of this kind of defer calls caudal delay. The method that reduces caudal delay currently makes a large number of CPU kernels in the server are in open position, the request that so that be handled quickly,passes into.

The computer crosses the data processing with kernel sensitive to time allocation newly, can answer an user quickly!

Rich division garden - scientific popular science: But this means kernel greater part to divide time to be in unused condition, and the server continues to use the sources of energy is to maintain power supply only. Data center can include 10 thousand tens of servers, although the efficiency of every server has a bit,rise accordingly, also can save millions dollar. Or, a few systems are according to working load between applied process reapportion kernel.

But this happens in millisecond above -- it is request of current fast rhythm about one thousandth what what want rate. Await too long the function that also can drop applied program, because be before distributive time,any information without processing won't send an user. In a paper that publishs on design of USENIX network system and implementation congress, researcher developed a kind of faster core dispatching system, the name is Shenango, it can reduce caudal delay, implementation is efficient at the same time.

Masschusetts Institute of Technology (a when MIT) researcher develops new system, allocate the data processing with sensitive time through crossing CPU kernel better, ensure hardware is efficient move, improved the efficiency that data center high speed operates. Picture: Massachusetts Institute Of Technology

Above all, a kind of new algorithm can detect what application process is processing data hardly. Next, software package allocates disengaged kernel to handle the working load that uses a program. Science of the first author, computer and artificial intelligence lab (the Amy Ousterhout of a doctoral student of CSAIL) says: Be in data center, balance exists between efficiency and defer, you need to comparing kernel of the reapportion on the granuality with every finer millisecond really. Shenango accessible for servicing implement " administration produces an operation inside very short time, execute these operations efficiently. Differ ladenly according to the size of data center and job, economic energy resources and cost also can differ somewhat. But overall objective is the CPU utilization rate that raises data center, make every core can get be usinged very well.

Current optimal CPU utilization rate is about 60% , but researcher expresses, their system may raise this one word 100% . The coauthor Adam Belay of the researcher of electronic project and the assistant professor with scientific computer, CSAIL, paper says: The utilization rate of data center is very low now, this is a very serious problem, a place that is in data center is inextricability. But this system is a crucial factor that raises utilization rate. The lecture that with what Ousterhout publishs a paper together with Belay still Fuji electrify enrages project and computer science to fasten teachs Hari Balakrishnan, the doctoral student Jonathan Behrens of CSAIL and Joshua Fried.

Efficient congest detects

Be in real data center, shenango algorithm and software will move on every server in data center, all server aux will be able to is enough each other correspond. The first innovation of this system is a kind of new congest check calculates a standard. The data that examination of every 5 small seconds queues up to wait to handle every application order wraps this algorithm. If data bag still is waiting to observe for the last time, this algorithm notices the delay that has 5 small seconds at least. It returns an examination to whether any computational progresses that call line Cheng wait to carry out. If so, the system thinks this is " crowded " applied process. Look very simple, but it is very important that the structure of alignment detects to congest of class of implementation small second. Traditional think of a way is the time stand sth on end that the data that lets software check every to queue up includes, this will spend too much time. Researcher is in call " annular buffer " alignment of the implementation in efficient structure, these structures can imagine the different chamfer on annulus.

The data bag of the first input enters an initiative chamfer. When new data arrives, they are put into the follow-up chamfer all round annulus. Normally, these structures are used at entering the data processing that goes out first first, collect data from initiative chamfer and work to end chamfer. However, the system of researcher stores data is wrapped momently only, till application the program can handle them. In the meantime, the data bag of memory can be used at congest examination.

This algorithm needs to compare alignment only medium two are nodded (the place that the first data includes and the last data include the place before 5 small seconds) whether did bag of OK and certain data encounter delay. Can observe this at 2 o'clock, every 5 small seconds dog their process, look how many data has been handled. Because the structure is very simple, every core needs to be done only. If you examine 24 nucleuses, can undertake 24 times checking inside 5 small seconds, this is very good.

Intelligent allocation

The 2nd innovation is IOKernel, it is the central software center that data bag guiding reachs to use a program appropriately. IOKernel still uses congest check to calculate a standard, can give congest application quickly the order kernel allocation, speed wants fast class of a few amounts. For example, IOKernel may see process of a certain application needs what small second treats rate to pass into data bag. If because lack a kernel,use a program and congest, IOKernel can stand to give application the order disengaged kernel allocation namely. If it still sees another application process is running not sensitive to time kernel, kernel of wh some of which of its meeting capture, give the application of congest the order their reapportion. The data after processing returns IOKernel to send answer.

What IOKernel pays close attention to is the kernel that what application process needs to do not have core, try to who found out to overload, need more nucleuses, offer a nucleus to them as soon as possible, such they won't lag behind, also won't have very big delay. The close together communication between process of IOKernel, algorithmic, application and server hardware " be unique in data center, allow Shenango to not have seam the ground to move, this system has global visibility to the thing that produces on every server. The content that it can see move in every the hardware that offers data package, kernel and every use the busy rate of the program. It accomplished this on the measure of small second. Next, researcher will undertake optimizing to Shenango, in order to realize real data center. For this, they ensure software can handle very high data handling capacity to have proper safe character.