CN110209350A - It is a kind of to mix in storage architecture HPC system using the dynamic dispatching method of I/O request - Google Patents

It is a kind of to mix in storage architecture HPC system using the dynamic dispatching method of I/O request Download PDF

Info

Publication number
CN110209350A
CN110209350A CN201910386909.8A CN201910386909A CN110209350A CN 110209350 A CN110209350 A CN 110209350A CN 201910386909 A CN201910386909 A CN 201910386909A CN 110209350 A CN110209350 A CN 110209350A
Authority
CN
China
Prior art keywords
application
request
hdd
random
ssd
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910386909.8A
Other languages
Chinese (zh)
Other versions
CN110209350B (en
Inventor
石宣化
金海�
杨莹
姜焰
花昱圣
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN201910386909.8A priority Critical patent/CN110209350B/en
Publication of CN110209350A publication Critical patent/CN110209350A/en
Application granted granted Critical
Publication of CN110209350B publication Critical patent/CN110209350B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601Interfaces specially adapted for storage systems
    • G06F3/0602Interfaces specially adapted for storage systems specifically adapted to achieve a particular effect
    • G06F3/061Improving I/O performance
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601Interfaces specially adapted for storage systems
    • G06F3/0628Interfaces specially adapted for storage systems making use of a particular technique
    • G06F3/0655Vertical data movement, i.e. input-output transfer; data movement between one or more hosts and one or more storage devices
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/06Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
    • G06F3/0601Interfaces specially adapted for storage systems
    • G06F3/0668Interfaces specially adapted for storage systems adopting a particular infrastructure
    • G06F3/0671In-line storage system
    • G06F3/0683Plurality of storage devices
    • G06F3/0685Hybrid storage combining heterogeneous device types, e.g. hierarchical storage, hybrid arrays

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Debugging And Monitoring (AREA)

Abstract

The invention discloses the dynamic dispatching methods in a kind of mixing storage architecture HPC system using I/O request, belong to High Performance Computing field.The present invention indicates the access module feature of application using random degree, dynamic dispatching I/O request, the SSD insensitive to random degree is written in the random more application of selection, the HDD sensitive to random degree is written in the random lesser application of degree, HDD handles the request of continuous mode as far as possible, SSD handles the request of stochastic model as far as possible, to mitigate I/O interference problem.It is distributed metastable phenomenon according to the I/O resource contention Time Bandwidth of discovery, is proposed for the calculation method of bandwidth allocation, can be used for predicting storing the dynamic load of equipment.The random degree of connected applications and can two parameters of obtainable bandwidth in storage equipment, utilize application access mode characteristic and different storage device load characteristic, rational management I/O request, preceding completion is run between ensureing application at runtime, system performance is improved simultaneously, to ensure application service quality.

Description

It is a kind of to mix in storage architecture HPC system using the dynamic dispatching method of I/O request
Technical field
The invention belongs to High Performance Computing field, more particularly, to being answered in a kind of mixing storage architecture HPC system The dynamic dispatching method requested with I/O.
Background technique
Multi-tenant multi-load scene is in cloud HPC (High Performance Computing, high-performance calculation) system It is increasingly common, it means that the storage resource in system is shared between more and more different users.So in object It manages in the limited situation of resource, is most important using service quality guarantee is provided with tenant.Service quality guarantee refers to protecting Card application performance be able to maintain that one reach an agreement on reasonable threshold value within the scope of, no matter the access request of other application arrival speed How rate, the random degree of request and succession and the read-write ratio of request etc. change, and all do not interfere with another application Service performance.The also scale calculated constantly expands, and needs more and more storage resources, while limited storage resource needs Serve more applications.When it is multiple application simultaneously access storage service when, they will compete I/O resource, this will lead to I/O gather The wide degradation of crossed belt.These have the I/O request of different access mode to mix, and current main storage is situated between Matter --- mechanical hard disk (Hard Disk Drive, HDD) can effectively handle the request of continuous mode, but random in processing When the request of mode, I/O can be caused to interfere because frequent positioning leads to serious magnetic head tracking expense.
In order to solve the serious degradation problem of random request bring, many solutions are suggested, wherein using new Type storage medium becomes preferable selection.Since solid state hard disk (Solid State Drive, SSD) is had excellent performance and to random Insensitive characteristic is requested, and as resident storage, every GB cost is far below DRAM, therefore SSD is passed through frequently as most common High-speed processing apparatus.And in HPC system, due to there is great storage demand, and the price of SSD is relatively high, so with SSD replaces HDD to be still unpractical completely.Therefore the mixing storage architecture of SSD and HDD is just at a preferable selection.
However, in current mixing storage architecture, mostly application competition I/O still very original for the scheduling of application I/O Resource can still occur I/O interference, cause application service quality that cannot ensure, performance of storage system it is low.
Summary of the invention
In view of the drawbacks of the prior art, it is an object of the invention to solve the I/O scheduling of prior art mixing storage architecture The technical problem that application service quality cannot ensure, performance of storage system is low.
To achieve the above object, in a first aspect, the embodiment of the invention provides answer in a kind of mixing storage architecture HPC system The dynamic dispatching method requested with I/O, method includes the following steps:
S1. all I/O request for the file system layer for reaching HPC system is grouped, the request from same application It is assigned to same group;
S2. judge whether N=1 is true, if so, this is write direct HDD using corresponding request;Otherwise, judge 1 < N≤ N0It is whether true, if so, calculating the random degree of each application and sorting according to random degree, by the smallest application of random degree HDD is written in corresponding request, and SSD is written in the corresponding request of others application;Otherwise, it calculates the random degree of each application and presses It sorts according to random degree, SSD is written into the corresponding request of random more M application, remaining N-M application is corresponding Request write-in HDD;
S3. if there is operation is added in the application newly to arrive, calculate and compare the B of the applicationi-SAnd Bi-HDIf Bi-SSD≤ Bi-HDD, HDD is written into using corresponding request in this, if Bi-SSD> Bi-HDD, further compare the application and operate in HDD The random maximum application of degree, by the corresponding request write-in SSD of degree the greater random in the two, by degree random in the two compared with HDD is written in the corresponding request of small person;
Wherein, Bi-SSD、Bi-HDDRespectively application operate in SSD may distribute acquisition bandwidth, using operate in HDD can The bandwidth obtained, M B can be distributedi-SSDStart to be less than Bi-HDDCritical value, N be parallel number of applications, N0For parallel application Quantity preset threshold.
Specifically, the calculation method of the random degree of application is as follows: being ranked up to the offset of I/O request;If I/O is asked Offset is asked to sort the distance between latter two request equal to request size, it is believed that the two requests are continuously that random factor is then Value is 0, otherwise it is assumed that the two requests are random requests, then value is 1 to random factor;Random journey is calculated with following formula Spend R:
Wherein, R_Factor is random factor, and K is this using the total distance number in corresponding all requests.
Specifically, parallel number of applications preset threshold N0Value range is [3,10].
Specifically, concurrently application may distribute the bandwidth B of acquisitioniCalculation formula is as follows:
Wherein, BiIndicate that i-th of application operates in storage equipment and may distribute the bandwidth of acquisition, B is indicated under storage equipment Aggregate bandwidth, PiIndicate the proportionality coefficient of i-th of application request size, QiIndicate i-th of application request size, n indicates parallel Number of applications.
Specifically, the aggregate bandwidth B under HDDHDDCalculation formula is as follows:
Wherein, BpeakFor the peak bandwidth of HDD under no I/O disturbed condition,To operate in HDD using random degree R's Mean value.
Specifically, the aggregate bandwidth B under SSDSSDFor the peak bandwidth of solid state hard disk.
Second aspect, the embodiment of the invention provides a kind of computer readable storage medium, the computer-readable storage mediums Computer program is stored in matter, which realizes that mixing described in above-mentioned first aspect stores when being executed by processor Using the dynamic dispatching method of I/O request in framework HPC system.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect Fruit:
1. the present invention indicates the access module feature of application using random degree, convenient for combining storage device characteristics, dynamic Dispatch I/O request.The SSD insensitive to random degree is written in the random more application of selection, and random degree is lesser to answer With the HDD sensitive to random degree is written, HDD handles the request of continuous mode as far as possible, and SSD handles stochastic model as far as possible Request, to alleviate I/O interference problem.
2. the present invention is distributed metastable phenomenon according to the I/O resource contention Time Bandwidth of discovery, it is proposed for bandwidth point The calculation method matched can be used for predicting storing the dynamic load of equipment, and in order to dynamic sensing load, it is suitable to facilitate selection Store equipment scheduling I/O request.
3. the random degree of connected applications of the present invention and in storage equipment can two parameters of obtainable bandwidth, using answering With access module characteristic and different storage device load characteristic, rational management I/O request ensures preceding completion between application at runtime Operation, while system performance is improved, to ensure application service quality.
Detailed description of the invention
Fig. 1 is the dynamic dispatching in a kind of mixing storage architecture HPC system provided in an embodiment of the present invention using I/O request Method flow diagram;
Fig. 2 is the dynamic dispatching schematic diagram that the embodiment of the present invention one provides;
Fig. 3 is dynamic dispatching schematic diagram provided by Embodiment 2 of the present invention;
Fig. 4 is the dynamic dispatching schematic diagram that the embodiment of the present invention three provides;
Fig. 5 (a) is the experimental result schematic diagram of IOR1 provided in an embodiment of the present invention;
Fig. 5 (b) is the experimental result schematic diagram of IOR2 provided in an embodiment of the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and It is not used in the restriction present invention.
As shown in Figure 1, using the dynamic dispatching method of I/O request, this method packet in a kind of mixing storage architecture HPC system Include following steps:
S1. all I/O request for the file system layer for reaching HPC system is grouped, the request from same application It is assigned to same group;
S2. judge whether N=1 is true, if so, this is write direct HDD using corresponding request;Otherwise, judge 1 < N≤ N0It is whether true, if so, calculating the random degree of each application and sorting according to random degree, by the smallest application of random degree HDD is written in corresponding request, and SSD is written in the corresponding request of others application;Otherwise, it calculates the random degree of each application and presses It sorts according to random degree, SSD is written into the corresponding request of random more M application, remaining N-M application is corresponding Request write-in HDD;
S3. if there is operation is added in the application newly to arrive, calculate and compare the B of the applicationi-SSDAnd Bi-HDIf Bi-SSD≤ Bi-HD, HDD is written into using corresponding request in this, if Bi-SSD> Bi-HDD, further compare the application and operate in HDD with SSD is written in the corresponding request of degree the greater random in the two by the maximum application of machine degree, and degree random in the two is smaller HDD is written in the corresponding request of person;
Wherein, Bi-SSD、Bi-HDDRespectively application operate in SSD may distribute acquisition bandwidth, using operate in HDD can The bandwidth obtained, M B can be distributedi-SSDStart to be less than Bi-HCritical value, N be parallel number of applications, N0For parallel application Quantity preset threshold.
Step S1. is grouped all I/O request for the file system layer for reaching HPC system, from same application Request is assigned to same group.
One application may correspond to multiple requests, set to guarantee that the request of the same application is written into same storage It is standby, all I/O request for the file system layer for reaching HPC system is grouped, the request from same application is assigned to together One group.
Step S2. judges whether N=1 is true, if so, this is write direct HDD using corresponding request;Otherwise, judge 1 <N≤N0It is whether true, if so, calculate the random degree of each application and sort according to random degree, random degree is the smallest HDD is written using corresponding request, SSD is written in the corresponding request of others application;Otherwise, the random degree of each application is calculated And sort according to random degree, SSD, remaining N-M application pair is written into the corresponding request of random more M application The request write-in HDD answered.
HDD is that random I/O request is very sensitive for access module, may result in frequent magnetic head tracking, thus Performance is caused to decline.SSD is insensitive for stochastic model, can efficiently handle these requests.So using random degree table The feature of existing application access mode, so that the suitable storage equipment of selection is respectively written into these requests.
According to it is initial when Parallel application quantity N it is different, be divided into the dynamic dispatching that three kinds of situations carry out I/O request.
This is write direct HDD using corresponding request by situation 1:N=1.
Situation 2:1 < N≤N0, it calculates the random degree of each application and sorts according to random degree, random degree is minimum Using corresponding request be written HDD, others application corresponding request write-in SSD.
Situation 3:N > N0, it calculates the random degree of each application and sorts according to random degree, it will random more M A to apply corresponding request that SSD is written, HDD, M B is written in the corresponding request of remaining N-M applicationi-SSDStart to be less than Bi-HDD Critical value, Bi-SSD、Bi-HDDRespectively application operates in SSD and may distribute the bandwidth of acquisition, may divide using operating in HDD Bandwidth with acquisition.
At this time if operation M are applied in SSD, N-M are applied in HDD, Bi-SSDJust start to be less than Bi-HDD, then M is to face Dividing value.N0Value range be [3,10], preferably 5.
The calculation method of the random degree of application is as follows: firstly, being ranked up to the offset of I/O request.Random factor (Random Factor) depends on I/O request and deviates the distance between latter two request of sorting.If it is big that distance is equal to request It is small, it is believed that the two requests are that continuously, then value is 0 to random factor.Otherwise it is assumed that the two requests are random requests, with Then value is 1 to the machine factor.Finally, random degree can be calculated with following formula.
Wherein, R_Factor is random factor, and K is this using the total distance number in corresponding all requests.
Operation is added if there is the application newly to arrive in step S3., calculates and compare the B of the applicationi-SSAnd Bi-HDDIf Bi-SS ≤Bi-HDD, HDD is written into using corresponding request in this, if Bi-SSD> Bi-HDD, further compare the application and operate in HDD The maximum application of random degree, SSD is written into the corresponding request of degree the greater random in the two, by degree random in the two HDD is written in the corresponding request of smaller.
Many experiments are run, reality is combined in different storage device to different access mode, applying for size of request It tests, analyzes experimental result.It is learnt according to the bandwidth distribution observation concurrently run to mostly application, when I/O request passes through file system When layer, the service that they are obtained will comply with relatively-stationary distribution, and the rule distribution almost divided equally is presented in bandwidth allocation.This meaning Within the given duration, can be evaluated whether to obtain the processing capacity that equipment is stored in competitive environment, i.e. processing is how many special Fixed I/O request.
Using this phenomenon, the present invention proposes a kind of calculate concurrently using the bandwidth B that may distribute acquisitioni, for characterizing The loading condition of equipment is stored in the process of running.
Wherein, BiIndicate that i-th of application operates in storage equipment and may distribute the bandwidth of acquisition, B is indicated under storage equipment Aggregate bandwidth, PiIndicate the proportionality coefficient of i-th of application request size, QiIndicate i-th of application request size, n indicates parallel Number of applications.According to the bandwidth distribution almost divided equally, PiValue is 1.QiValue 64K, 128K, 256K etc..
For example, one is 64k, and one is 128k, then requesting size is answering for 64k if there are two application while running The bandwidth of 1/3B is obtained with request, request size is that the application request of 128k obtains the bandwidth of 2/3B.
It is found according to experiment, in HDD, the random degree of I/O bandwidth and application request is in the relationship of approximately linear, with The increase of the random degree of application request, I/O bandwidth are gradually reduced.So can according to mechanical hard disk mixing storage architecture in, The random degree of peak bandwidth and application request without HDD under I/O disturbed condition, calculates in mechanical hard disk in the storage architecture Under concurrently run using total bandwidth B obtainedHDD, so as to further calculate out while operating in an application in HDD The bandwidth B obtained may be distributedi-HDD
It is as follows to calculate HDD aggregate bandwidth formula:
Wherein, BpeakFor the peak bandwidth of HDD under no I/O disturbed condition,To operate in HDD using random degree Mean value.
It is insensitive for random request due to solid-state hard disk SSD self-characteristic (almost 0 access delay), it is believed that mixed Close the aggregate bandwidth B of solid state hard disk under storage architectureSSDIt is same so as to further calculate out for the peak bandwidth of solid state hard disk When operate in an application in SSD and may distribute the bandwidth B of acquisitioni-SSD
As shown in Fig. 2, only one in embodiment one applies isolated operation in mixing storage architecture HPC system, and do not have There is other application while running, using not interfered by I/O, the end of run time of the application can be met by write-in HDD, So HDD is write direct in selection when one application of operation.
As shown in figure 3, in embodiment two two application simultaneously run, calculate first the two apply random degree, so Compare the random degree of the two applications afterwards.The random lesser application of degree is written into HDD, another is then written into SSD.
There is N number of application to start simultaneously at operation in embodiment three, calculates application first and operate in SSD and HDD and may distribute Obtain bandwidth Bi-SSDAnd Bi-HDD;Then a critical value is found, at this time if operation M are applied in SSD, N-M are applied HDD, Bi-SSDJust start to be less than Bi-HDD;Secondly, calculating the random degree of N number of application, and it is ranked up, M is a random before selecting More application operates in SSD, and remaining N-M application then operates in HDD.As shown in figure 4, original N number of application It is currently running, continuous at the same time to have new application that operation is added again, calculating a apply of N+1 first may in SSD and HDD The bandwidth B being assigned toi-SAnd Bi-HDDIf Bi-SSDLess than Bi-HDD, illustrate that being not suitable for again more operations N+1 in SSD at this time answers With, if selection operates in SSD, considerable bandwidth will not be not only obtained, the application operated in SSD can be also tied down, So the N+1 application is write direct in HDD;If Bi-SSDGreater than Bi-HDD, the random degree of the N+1 application is calculated, than Compared with the random degree of the maximum application of random degree run in the N+1 application and HDD, if the N+1 apply it is random Degree in HDD less than or equal to what is applied, then the application originally run in HDD will be redirected to SSD, and N+1 exclusive HDD, on the contrary it will be in the N+1 application write-in SSD.
Mean that limited physical resource is shared by more and more users in cloud HPC environment, the present invention does not chase after blindly Peak performance is sought, but distributes I/O resource according to application demand to ensure the service quality of application, similar to pressing in cloud computing It needs to service.The present invention defines the demand of application using the end of run time, and the end of run time refers to applying and individually transport Runing time when row is interference-free.Each application has the end of run time, when implemented between be more than end of run when, just very It is difficult to guarantee the service quality of application.So at its best, each application can oneself the end of run time it Before be finished.Meanwhile application the end of run time and handling capacity can be used as the judgment basis of service quality guarantee.Cause This, the invention proposes a kind of Dynamic Scheduling Strategies, and being able to satisfy for each application is ensured using the characteristic of different storage device The end of run time, to improve bandwidth availability ratio and provide high I/O performance.
The present invention carries out being configured with 10 nodes in the cluster of confirmatory experiment, wherein and 8 nodes are as calculate node, and 2 Node is as I/O node, wherein the memory size of calculate node is configured to 64GB, and configures the mechanical disk of 300GB, I/O section The memory size of point is configured to 8GB, and configures the solid state hard disk of 300GB, is connected to the network between node using Infiniband, remaining It configures identical.In each calculate node, the MPICH compiled by ROMIO is mounted with to realize Parallel I/O, and DDL-QoS Realization in used MPI-IO to handle MPICH, while OrangeFS (original file system) is mounted on two I/O nodes With DDL-QoS (being configured with dynamic dispatching method proposed by the present invention).In an experiment, main benchmark to be used is IOR, this is main Flow a kind of parallel file system commonly testing tool.Experiment simultaneously operation two IOR examples, according between two examples when Between be spaced different, carried out five groups of experiments respectively.Time interval between example is 0,9,18,27,36 second.Each IOR example The data of 8GB are written, each I/O request size is 512KB, and default process number is 32.
Compare Fig. 5 (a) and Fig. 5 (b) it is found that the performance of IOR1 improves 2.5 in DDL-QoS when time interval is 0s Times, the performance of IOR2 improves 3.79 times.This is because the I/O between two IOR examples interferes journey when time interval is 0 Degree is maximum, leads to performance degradation.The DDL-QoS that the present invention designs can analyze the access module of each application program, IOR1 is the access module of Segmented-Contiguous a kind of, and good performance still can be obtained on HDD, so IOR1 is still written into HDD.When IOR2 is in the access module of Segmented-Random, need to be written into asking at random Insensitive SSD is sought, to obtain maximum performance improvement.
When the time, which is spaced apart, begins to increase, the IOR1 of first operation can during this period of time monopolize the I/O money of server Source.When parallel operation is added in IOR2, it starts to generate I/O interference.Therefore, in original OrangeFS, the performance of IOR1 It is gradually increased with the increase of time interval.In experiment of the invention, when time interval is 36 seconds, in OrangeFS The performance of IOR1 maximizes, this is identical as the performance in DDL-QoS, because absolutely not I/O is interfered IOR1 during operation. In IOR2, time interval is bigger, and the I/O interference generated between IOR1 and IOR2 is smaller.Equally, with the increasing of time interval Add, the performance of IOR2 can also improve in OrangeFS.However, in DDL-QoS, the performance of IOR2 with time interval increase And it reduces.This is because needing to be written SSD can be reduced to avoid the part that I/O is interfered with the reduction of I/O annoyance level.It writes The data volume for entering HDD also increased, finally suitable with the performance for being separately written HDD.
More than, the only preferable specific embodiment of the application, but the protection scope of the application is not limited thereto, and it is any Within the technical scope of the present application, any changes or substitutions that can be easily thought of by those familiar with the art, all answers Cover within the scope of protection of this application.Therefore, the protection scope of the application should be subject to the protection scope in claims.

Claims (7)

1. using the dynamic dispatching method of I/O request in a kind of mixing storage architecture HPC system, which is characterized in that this method packet Include following steps:
S1. all I/O request for the file system layer for reaching HPC system is grouped, the request from same application is divided To same group;
S2. judge whether N=1 is true, if so, this is write direct HDD using corresponding request;Otherwise, judge 1 < N≤N0It is No establishment, if so, calculate the random degree of each application and sort according to random degree, the smallest application of random degree is corresponding Request be written HDD, others application corresponding request write-in SSD;Otherwise, calculate the random degree of each application and according to SSD is written in the corresponding request of random more M application by the sequence of machine degree, and remaining N-M is applied corresponding request HDD is written;
S3. if there is operation is added in the application newly to arrive, calculate and compare the B of the applicationi-SSDAnd Bi-HDDIf Bi-SSD≤Bi-HDD, HDD is written into using corresponding request in this, if Bi-SSD> Bi-HDD, further compare the application and operate in the random journey in HDD Maximum application is spent, SSD is written into the corresponding request of degree the greater random in the two, by degree smaller couple random in the two The request write-in HDD answered;
Wherein, Bi-SSD、Bi-HDDRespectively application operates in SSD and may distribute the bandwidth of acquisition, may distribute using operating in HDD The bandwidth of acquisition, M Bi-SSDStart to be less than Bi-HDDCritical value, N be parallel number of applications, N0It is pre- for parallel number of applications If threshold value.
2. dynamic dispatching method as described in claim 1, which is characterized in that the calculation method of the random degree of application is as follows: The offset of I/O request is ranked up;If the distance between latter two request of I/O request offset sequence is equal to request size, Think the two requests be it is continuous, then value is 0 to random factor, otherwise it is assumed that the two requests are random requests, it is random because Sub then value is 1;Random degree R is calculated with following formula:
Wherein, R_Factor is random factor, and K is this using the total distance number in corresponding all requests.
3. dynamic dispatching method as claimed in claim 1 or 2, which is characterized in that parallel number of applications preset threshold N0Value Range is [3,10].
4. dynamic dispatching method as described in any one of claims 1 to 3, which is characterized in that concurrently application may distribute acquisition Bandwidth BiCalculation formula is as follows:
Wherein, BiIndicate that i-th of application operates in storage equipment and may distribute the bandwidth of acquisition, B indicates the polymerization under storage equipment Bandwidth, PiIndicate the proportionality coefficient of i-th of application request size, QiIndicate i-th of application request size, n indicates Parallel application number Amount.
5. dynamic dispatching method as claimed in claim 4, which is characterized in that the aggregate bandwidth B under HDDHDDCalculation formula is such as Under:
Wherein, BpeakFor the peak bandwidth of HDD under no I/O disturbed condition,To operate in HDD using the equal of random degree R Value.
6. dynamic dispatching method as claimed in claim 4, which is characterized in that the aggregate bandwidth B under SSDSSDFor solid state hard disk Peak bandwidth.
7. a kind of computer readable storage medium, which is characterized in that be stored with computer on the computer readable storage medium Program realizes such as mixing storage architecture as claimed in any one of claims 1 to 6 when the computer program is executed by processor Using the dynamic dispatching method of I/O request in HPC system.
CN201910386909.8A 2019-05-10 2019-05-10 Dynamic scheduling method for application I/O (input/output) request in HPC (high performance computing) system of hybrid storage architecture Active CN110209350B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910386909.8A CN110209350B (en) 2019-05-10 2019-05-10 Dynamic scheduling method for application I/O (input/output) request in HPC (high performance computing) system of hybrid storage architecture

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910386909.8A CN110209350B (en) 2019-05-10 2019-05-10 Dynamic scheduling method for application I/O (input/output) request in HPC (high performance computing) system of hybrid storage architecture

Publications (2)

Publication Number Publication Date
CN110209350A true CN110209350A (en) 2019-09-06
CN110209350B CN110209350B (en) 2020-07-10

Family

ID=67785638

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910386909.8A Active CN110209350B (en) 2019-05-10 2019-05-10 Dynamic scheduling method for application I/O (input/output) request in HPC (high performance computing) system of hybrid storage architecture

Country Status (1)

Country Link
CN (1) CN110209350B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111679813A (en) * 2020-08-11 2020-09-18 南京云联数科科技有限公司 Method for information processing, electronic device, and storage medium
CN114691698A (en) * 2022-04-24 2022-07-01 北京梦蓝杉科技有限公司 Data processing system and method for computer system

Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130159597A1 (en) * 2011-12-19 2013-06-20 Electronics And Telecommunications Research Institute Hybrid storage device and method of operating the same
CN104216988A (en) * 2014-09-04 2014-12-17 天津大学 SSD (Solid State Disk) and HDD(Hard Driver Disk)hybrid storage method for distributed big data
CN104376094A (en) * 2014-11-24 2015-02-25 浪潮电子信息产业股份有限公司 File hierarchical storage method and system taking visit randomness into consideration
US20160179683A1 (en) * 2014-12-23 2016-06-23 Prophetstor Data Services, Inc. Ssd caching system for hybrid storage
US20160299697A1 (en) * 2015-04-08 2016-10-13 Prophetstor Data Services, Inc. Workload-aware i/o scheduler in software-defined hybrid storage system
CN106469193A (en) * 2016-08-30 2017-03-01 北京航空航天大学 Multi load metadata I/O service quality performance support method and system
US20170060472A1 (en) * 2015-08-31 2017-03-02 Linkedin Corporation Transparent hybrid data storage
CN107015763A (en) * 2017-03-03 2017-08-04 北京中存超为科技有限公司 Mix SSD management methods and device in storage system
CN107193487A (en) * 2017-04-26 2017-09-22 华中科技大学 A kind of random access recognition methods and system
CN107844269A (en) * 2017-10-17 2018-03-27 华中科技大学 A kind of layering mixing storage system and method based on uniformity Hash
US20180307413A1 (en) * 2013-11-27 2018-10-25 Alibaba Group Holding Limited Control of storage of data in a hybrid storage system
CN109144411A (en) * 2018-07-24 2019-01-04 中国电子科技集团公司第三十八研究所 Data center's hybrid magnetic disc array and its data dynamic migration strategy
CN109240611A (en) * 2018-08-28 2019-01-18 郑州云海信息技术有限公司 The cold and hot data hierarchy method of small documents, small documents data access method and its device
CN109697033A (en) * 2018-12-19 2019-04-30 中国人民解放军国防科技大学 Tile record disk sensing storage caching method and system

Patent Citations (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20130159597A1 (en) * 2011-12-19 2013-06-20 Electronics And Telecommunications Research Institute Hybrid storage device and method of operating the same
US20180307413A1 (en) * 2013-11-27 2018-10-25 Alibaba Group Holding Limited Control of storage of data in a hybrid storage system
CN104216988A (en) * 2014-09-04 2014-12-17 天津大学 SSD (Solid State Disk) and HDD(Hard Driver Disk)hybrid storage method for distributed big data
CN104376094A (en) * 2014-11-24 2015-02-25 浪潮电子信息产业股份有限公司 File hierarchical storage method and system taking visit randomness into consideration
US20160179683A1 (en) * 2014-12-23 2016-06-23 Prophetstor Data Services, Inc. Ssd caching system for hybrid storage
US20160299697A1 (en) * 2015-04-08 2016-10-13 Prophetstor Data Services, Inc. Workload-aware i/o scheduler in software-defined hybrid storage system
US20170060472A1 (en) * 2015-08-31 2017-03-02 Linkedin Corporation Transparent hybrid data storage
CN106469193A (en) * 2016-08-30 2017-03-01 北京航空航天大学 Multi load metadata I/O service quality performance support method and system
CN107015763A (en) * 2017-03-03 2017-08-04 北京中存超为科技有限公司 Mix SSD management methods and device in storage system
CN107193487A (en) * 2017-04-26 2017-09-22 华中科技大学 A kind of random access recognition methods and system
CN107844269A (en) * 2017-10-17 2018-03-27 华中科技大学 A kind of layering mixing storage system and method based on uniformity Hash
CN109144411A (en) * 2018-07-24 2019-01-04 中国电子科技集团公司第三十八研究所 Data center's hybrid magnetic disc array and its data dynamic migration strategy
CN109240611A (en) * 2018-08-28 2019-01-18 郑州云海信息技术有限公司 The cold and hot data hierarchy method of small documents, small documents data access method and its device
CN109697033A (en) * 2018-12-19 2019-04-30 中国人民解放军国防科技大学 Tile record disk sensing storage caching method and system

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
MING LI等: "《SSDUP:An Efficient SSD Write Buffer Using Pipeline》", 《2016 IEEE INTERNATIONAL CONFERENCE ON CLUSTER COMPUTING》 *
SHI XUANHUA: "《https://dl.acm.org/doi/pdf/10.1145/3079079.3079087?download=true》", 16 June 2017 *
祝青: "混合存储系统研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111679813A (en) * 2020-08-11 2020-09-18 南京云联数科科技有限公司 Method for information processing, electronic device, and storage medium
CN111679813B (en) * 2020-08-11 2020-11-06 南京云联数科科技有限公司 Method for information processing, electronic device, and storage medium
CN114691698A (en) * 2022-04-24 2022-07-01 北京梦蓝杉科技有限公司 Data processing system and method for computer system

Also Published As

Publication number Publication date
CN110209350B (en) 2020-07-10

Similar Documents

Publication Publication Date Title
CN105205014B (en) A kind of date storage method and device
US11556541B2 (en) Data query method, apparatus and device
CN107124472A (en) Load-balancing method and device, computer-readable recording medium
US7418568B2 (en) Memory management technique
US11093410B2 (en) Cache management method, storage system and computer program product
CN106406987A (en) Task execution method and apparatus in cluster
CN103327072A (en) Method for cluster load balance and system thereof
CN107526542A (en) Object storage device and its operating method
US20190332328A1 (en) Storage Controller and IO Request Processing Method
CN106569892A (en) Resource scheduling method and device
CN104102646B (en) The method, apparatus and system of data processing
CN110209350A (en) It is a kind of to mix in storage architecture HPC system using the dynamic dispatching method of I/O request
CN110636388A (en) Service request distribution method, system, electronic equipment and storage medium
WO2021259246A1 (en) Resource scheduling method and apparatus, electronic device, and computer-readable storage medium
WO2021258512A1 (en) Data aggregation processing apparatus and method, and storage medium
US20190149478A1 (en) Systems and methods for allocating shared resources in multi-tenant environments
CN109756429A (en) Bandwidth allocation methods and equipment
CN109788013B (en) Method, device and equipment for distributing operation resources in distributed system
US11194619B2 (en) Information processing system and non-transitory computer readable medium storing program for multitenant service
CN107800744A (en) A kind of service request retransmission method, apparatus and system
WO2018205890A1 (en) Task assignment method and system of distributed system, computer readable storage medium and computer device therefor
Gao et al. A load-aware data migration scheme for distributed surveillance video processing with hybrid storage architecture
CN112667392B (en) Cloud computing resource allocation method and device, computer equipment and storage medium
US20130151794A1 (en) Memory controller and memory control method
US9894670B1 (en) Implementing adaptive resource allocation for network devices

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant