CN1971562A - Distributing method of object faced to object storage system - Google Patents

Distributing method of object faced to object storage system Download PDF

Info

Publication number
CN1971562A
CN1971562A CN 200610125189 CN200610125189A CN1971562A CN 1971562 A CN1971562 A CN 1971562A CN 200610125189 CN200610125189 CN 200610125189 CN 200610125189 A CN200610125189 A CN 200610125189A CN 1971562 A CN1971562 A CN 1971562A
Authority
CN
China
Prior art keywords
file
burst
equipment
value
storage system
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.)
Pending
Application number
CN 200610125189
Other languages
Chinese (zh)
Inventor
王芳
冯丹
庞丽萍
刘群
潘磊颖
张顺达
吕松
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 CN 200610125189 priority Critical patent/CN1971562A/en
Publication of CN1971562A publication Critical patent/CN1971562A/en
Pending legal-status Critical Current

Links

Images

Landscapes

  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The objects distribution method facing the objects storage system is disclosed that belongs to the computer storage system facing objects with the characteristic of distributed storage system, the aim is combined with the advantages of Hasi algorism and the cutted piece algorism to improve the load balance of equipment, the concurrency and the reliability. The steps are: (1) determining the boundary value L of the file length; (2) obtaining the total number N of equipments when the original data sever is initialized; (3) obtaining the file length from clients, and the file length is compared with the defined boundary value L of the file length to judge the file is big or small; (4) if the result is small file, the object is mapped to the device by Hasi directly; (5)if the result is big file, the file will be split into several objects and located in the different devices. The invention reduces the overhead cost when the system with a mount of devices distributes the objects, and improves the concurrency and the reliability of devices. The invention can be used to distribute the object on one or more storage devices.

Description

Object distribution method in the object-oriented storage system
Technical field
The invention belongs to object oriented calculation machine storage system, be specifically related to the object distribution method in a kind of object-oriented storage system, be used for object distribution to one or more object storage equipment with distributed storage architecture characteristics.
Background technology
OO storage system is considered as file the set of object.These object distribution are in object storage equipment OSD (Object-Based Storage Device).Object is placed into and can makes system have higher capacity, handling capacity, reliability and extensibility in the different equipment.For how improving in the object-oriented storage system object to place Study on Efficiency less relatively.The object distribution method begins just to have determined the performance of system at the communication process of total system, also affects the load balancing of equipment room and the parallel efficiency of equipment room simultaneously.
The object-oriented storage system is made up of client, meta data server and memory device three parts.This tripartite communication process can be described as: client sends the reading and writing of files request to meta data server, the object information of meta data server backspace file correspondence, and client is visited corresponding apparatus according to object information.In some distributed file systems, data are stored on the smart machine that can directly visit by network, and the metadata of data of description information is by special meta data server management.The position of meta data server in the object-oriented storage system is extremely important, and one of its task becomes one or more objects with File mapping exactly and is stored on the equipment, so whether the method that distributes efficiently directly influences system performance.
The algorithm that existing object distribution method is used mainly contains two kinds.First kind is called hash algorithm, and the use hash function is mapped as an object with a file integral body and is placed in the equipment.Second kind is called the burst mapping algorithm, with the file mean allocation in each equipment.Hash algorithm and burst mapping algorithm all have relative merits separately.Hash algorithm use hash function with the file data approximate random be distributed on the different equipment, can realize the efficient of load balancing and higher distributed data.But this man-to-man mapping model can not make full use of the concurrency of equipment for big file; And hash algorithm on can expanding also existing problems because the result of hash function will be subjected to the influence of equipment sum, so hash algorithm supposes that usually the memory device sum is constant.The burst mapping algorithm has utilized the concurrency of each equipment.It has two advantages: the first, simplified the operation of client, and client does not need concrete distribution method; The second, change equipment sum does not influence existing map information on the meta data server, is with good expansibility.But when file hour, use concurrency some waste that seems, because the expense that connects to each equipment under the small documents situation has surpassed the parallel benefit of bringing of equipment, particularly when the equipment sum is bigger, file is divided into a lot of small objects, is to create object or connect all very consuming timely, and reliability also can reduce, because if an equipment goes wrong, total system will turn round.
Summary of the invention
The present invention proposes the object distribution method in a kind of object-oriented storage system, purpose is the advantage in conjunction with hash algorithm and burst mapping algorithm, avoids their shortcoming simultaneously as far as possible, improves the load balancing of equipment, the concurrency of equipment room and equipment dependability, thus system performance improved.
Object distribution method in a kind of object-oriented storage system of the present invention, order comprises the steps:
(1) determine file size boundary value L, L is 256KB, 512KB or 1MB;
(2) obtain equipment sum N when the meta data server initialization, N is a positive integer;
(3) obtain file size from client, the file size boundary value L that defines in file size and the algorithm is compared, judgement is big file or small documents;
(4), directly be hashed into an object and be mapped in the equipment if small documents;
(5) if big file becomes several objects with its burst, be placed on respectively in the different equipment.
Object distribution method in the described object-oriented storage system is characterized in that: described big file fragmentation step is comprised following process:
(1) determines the burst reference value n of all lengths file Reference
(2) according to equipment sum N and burst reference value n ReferenceJudge the file fragmentation number n p
Burst reference value n ReferenceGreater than equipment sum N, then burst is counted n The resultBe equipment sum N; Burst reference value n ReferenceLess than equipment sum N, then burst is counted n pFor burst is counted reference value n Reference
(3) with equipment by the load parameter value of being obtained from the memory device end, sort from low to high, choose the low preceding n of load parameter value pIndividual equipment;
(4) the even burst of file is mapped to the n that chooses pOn the individual equipment.
Object distribution method in the described object-oriented storage system is characterized in that: when (1) determines file size boundary value L, and should be according to the file proportion of all lengths in the concrete environment; (2) determine burst reference value n under the various file sizes ReferenceThe time, according to the situation of change that the cost performance of all lengths file fragmentation parallel transmission is counted n with burst, choose and make cost performance reach optimum burst number as burst reference value n Reference
Object distribution method in the described object-oriented storage system is characterized in that: determine that described cost performance reaches optimum burst reference value n ReferenceThe time, find first n value that satisfies with lower inequality to be n Reference:
Figure A20061012518900061
In the formula, a is the transmit leg expense, and b is the transmission time, and ε is the expectation value that increases the performance gain amplitude along with the burst number, and the ε reference value is 0.7 * 10 -3,
Figure A20061012518900062
If in the interval [1, n Theoretical] between do not satisfy the n of inequality, n then Reference=n Theoretical
Among the present invention, the file size boundary value determines according to the file proportion of all lengths in the concrete applied environment, and the file size of proportion maximum is decided to be the file size boundary value, and fundamental purpose is to bring into play the advantage of equipment end file system.
Network environment is meant the transmit leg expense and the network bandwidth; The theoretical best burst number of the file of all lengths Can obtain by following analysis:
The number of object storage equipment (OSD) and the relation between the degree of parallelism are quite complicated.Equipment transmission channel more at most is many more, helps parallel transmission.But there is certain influence the institute's time spent that connects to the entire system performance, and especially when linking number is a lot, this influence will make performance reduce greatly.Data sync also is a time-consuming job.If a lot of the corresponding same big files of object are assembled into file again with these file fragments and also will spend many CPU time.Therefore, when same file corresponding equipment number increased, the transmission degree of parallelism increased, and has also consumed extra system resource simultaneously.This relation can be described with following formula:
T p T = n × a + 1 n × b + δ ( n ) a + b - - - ( 1 )
Tp: the overall delay of parallel transmission file.
T: the overall delay of serial transmission file.
N: the corresponding number of devices that distributes of file.
A: transmit leg expense.
B: transmission time.
δ (t): the coordinated time between object, comprise synchronously, verification etc., be the function of n
The molecule of formula (1) comprises three part: n * a, b/n and δ (t).The transmit leg expense that n * a represents to connect n equipment be connect individual equipment n doubly.And b/n shows that the bandwidth of n equipment parallel transmission is n a times of individual equipment.Though δ (t) is an independent variable is the function of n,, can think approx that δ (t) is a constant because its variation with n is very little.So substitute δ (t) with constant c, formula (1) abbreviation also further is deformed into:
T p T = n × a a + b + 1 n × b a + b + c a + b - - - ( 2 )
T/Tp is the speed-up ratio of parallel transmission with respect to serial transmission, the speed-up ratio maximum just is equivalent to makes Tp/T minimum in the formula (2).And the method that the present invention proposes is mainly paid close attention to is to be a plurality of objects with a File mapping how, and promptly n is to the influence of Tp/T, so as long as find best n just can obtain maximum speed-up ratio.Order F ( n ) = n × a a + b + 1 n × b a + b , To its differentiate dF ( n ) dn = 0 , We can obtain
Figure A20061012518900083
Because d 2 F ( n ) dn 2 > 0 , F (n) exists
Figure A20061012518900085
There is minimum value at the place.Promptly work as Time transmission speed-up ratio maximum.If transmission time b is very big or send expense a when very little, it is very big that n will become.This also just shows if the transmission time is very little, then there is no need the parallel transmission object.
Wherein transmit leg expense a is determined by network condition, and transmission time b is by file size and network bandwidth decision, b=file size/network bandwidth.So under the condition of determining network environment, can basis The best burst number of pairing theory when obtaining the file optimal transmission performance of all lengths.In actual applications, make systemic price ratio reach the highest, also should consider increases the cost expense that number of devices is brought, and then makes cost performance reach optimum burst reference value n reference and is determined by following analysis:
Order F ( n ) = n × a a + b + 1 n × b a + b , Its differentiate is got dF ( n ) dn = a a + b - 1 n 2 × b a + b .
ε is the expectation value that increases the performance gain amplitude along with the burst number, and reference value is 0.7 * 10 -3Find first n value that satisfies inequality (3) to be n ReferenceIf, in the interval [1, n Theoretical] between do not satisfy the n of inequality (3), n then Reference=n Theoreticaln ReferenceThe physical meaning of substitution (3) formula represents that working as the burst number is n ReferenceThe time, Tp/T changes not too obvious with the burst number, can think that this moment, systemic price ratio reached the highest.
The present invention has reduced the expense of distributing object in the system that has large number quipments, has improved the concurrency and the reliability of equipment.
Description of drawings
Fig. 1 is a FB(flow block) of the present invention;
Fig. 2 is the FB(flow block) of the big file fragmentation of the present invention;
Fig. 3 is the graph of a relation of time ratio Tp/T and object number n;
Fig. 4 is transmission overall delay (a) 16-OSDs of algorithm of the present invention in each virtual system, (b) 32-OSDs, (c) 64-OSDs;
Fig. 5 is comparison (a) 16-OSDs of transmission overall delay in each virtual system of algorithm of the present invention and hash algorithm and burst mapping algorithm, (b) 32-OSDs, (c) 64-OSDs.
Embodiment
The present invention is described in further detail below in conjunction with embodiment.
The boundary value L of file size is decided to be 512KB: by statistical study, in the load of LLNL, it is 512KB that 85% object length is arranged approximately, 15% object length is less than 512KB, in design, the boundary value of file size is defined as 512KB thus, more helps bringing into play the advantage of equipment end file system.
In big file, chosen the situation of the most normal appearance: 1MB, 2MB, 4MB, 8MB, 16MB, 32MB, 64MB, 128MB, 256MB, 512MB and 1GB.Choose a kind of typical case on network: bandwidth is 1000Mbits/ second, and the expense factor is 80 microseconds.For the file of 1MB, Message size=8Mbits, Bandwidth=1000Mbits, b=1*8/1000=0.008s=8000 μ s, a=80 μ s, n = b / a = 10 .
As shown in Figure 3, for the file of 1MB and 2MB, the change of Tp/T is not obvious when n＞=10.For 4MB, the file of 8MB and 16MB, the variation of Tp/T is not obvious when n＞=20.Equally, to the file greater than 32MB, n＞=40 an o'clock Tp/T has changed with n hardly.Can obtain as drawing a conclusion: n Reference=10 (file size is 1MB-2MB), n Reference=20 (file size is 4MB-16MB), n Reference=40 (file size is 32MB-1GB).
Experimental situation is the PC of a 2.4GHz Intel Celeron CPU 512MB RAM, and operating system is Red Hat Linux 9.0, kernel version 2 .4.20.Use Matlab as simulation software.At first generate the array at random of an expression file size according to exponential distribution, object allocation algorithm of the present invention then is applied to calculate corresponding propagation delay time in these files.Algorithm is realized with the form of the M file of Matlab.Suppose that the Overhead expense factor is 80 microseconds, the network bandwidth is 1000Mb/s.And should be able to be faster than the equipment of selecting at random through selecting equipment after the ordering.According to the difference of OSD number, the situation of emulation is divided into following three classes: the system of 16 OSD, the system of the system of 32 OSD and 64 OSD.
Every statistics of table 1. object allocation algorithm
Number of devices Average (μ s) Standard deviation (μ s) Maximal value (μ s) Minimum value (μ s)
16 2.229474e+06 3.303443e+04 2.309076e+06 2.132241e+06
32 2.230677e+06 3.514477e+04 2.315964e+06 2.152755e+06
64 2.235837e+06 3.870051e+04 2.311283e+06 2.160869e+06
Every statistics of table 2. hash algorithm
Number of devices Average (μ s) Standard deviation (μ s) Maximal value (μ s) Minimum value (μ s)
16 4.274640e+06 1.382437e+05 4.644530e+06 3.883163e+06
32 4.255894e+06 1.180060e+05 4.511304e+06 4.010331e+06
64 4.187441e+06 1.448126e+05 4.512065e+06 3.857762e+06
Every statistics of table 3. slicing algorithm
Number of devices Average (μ s) Standard deviation (μ s) Maximal value (μ s) Minimum value (μ s)
16 1.572765e+06 8.640234e+03 1.595883e+06 1.548298e+06
32 2.751877e+06 3.687687e+03 2.759858e+06 2.744203e+06
64 5.308309e+06 2.262698e+03 5.313381e+06 5.303158e+06
From table numerical value and Fig. 5 as can be seen, the standard deviation of the object allocation algorithm overall delay among the present invention is all littler than hash algorithm and burst mapping algorithm.See that on the whole object allocation algorithm of the present invention is more stable than other two kinds of algorithms, and behaves oneself best when the object-oriented storage system has a large amount of OSD.

Claims (4)

1. the object distribution method in the object-oriented storage system, order comprises the steps:
(1) determine file size boundary value L, L is 256KB, 512KB or 1MB;
(2) obtain equipment sum N when the meta data server initialization, N is a positive integer;
(3) obtain file size from client, the file size boundary value L that defines in file size and the algorithm is compared, judgement is big file or small documents;
(4), directly be hashed into an object and be mapped in the equipment if small documents;
(5) if big file becomes several objects with its burst, be placed on respectively in the different equipment.
2. the object distribution method in the object-oriented storage system as claimed in claim 1 is characterized in that: described big file fragmentation step is comprised following process:
(1) determines the burst reference value n of all lengths file Reference
(2) according to equipment sum N and burst reference value n ReferenceJudge the file fragmentation number n p
Burst reference value n ReferenceGreater than equipment sum N, then burst is counted n The resultBe equipment sum N; Burst reference value n ReferenceLess than equipment sum N, then burst is counted n pFor burst is counted reference value n Reference
(3) with equipment by the load parameter value of being obtained from the memory device end, sort from low to high, choose the low preceding n of load parameter value pIndividual equipment;
(4) the even burst of file is mapped to the n that chooses pOn the individual equipment.
3. the object distribution method in the object-oriented storage system as claimed in claim 1 or 2 is characterized in that: when (1) determines file size boundary value L, and should be according to the file proportion of all lengths in the concrete environment; (2) determine burst reference value n under the various file sizes ReferenceThe time, according to the situation of change that the cost performance of all lengths file fragmentation parallel transmission is counted n with burst, choose and make cost performance reach optimum burst number as burst reference value n Reference
4. the object distribution method in the object-oriented storage system as claimed in claim 3 is characterized in that: determine that described cost performance reaches optimum burst reference value n ReferenceThe time, find first n value that satisfies with lower inequality to be n Reference:
| a a + b - 1 n 2 × b a + b | ≤ ϵ , (n=1,2 ..., n Theoretical)
In the formula, a is the transmit leg expense, and b is the transmission time, and ε is the expectation value that increases the performance gain amplitude along with the burst number, and the ε reference value is 0.7 * 10 -3, If in the interval [1, n Theoretical] between do not satisfy the n of inequality, n then Reference=n Theoretical
CN 200610125189 2006-11-29 2006-11-29 Distributing method of object faced to object storage system Pending CN1971562A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 200610125189 CN1971562A (en) 2006-11-29 2006-11-29 Distributing method of object faced to object storage system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 200610125189 CN1971562A (en) 2006-11-29 2006-11-29 Distributing method of object faced to object storage system

Publications (1)

Publication Number Publication Date
CN1971562A true CN1971562A (en) 2007-05-30

Family

ID=38112385

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 200610125189 Pending CN1971562A (en) 2006-11-29 2006-11-29 Distributing method of object faced to object storage system

Country Status (1)

Country Link
CN (1) CN1971562A (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101494658A (en) * 2008-01-24 2009-07-29 华为技术有限公司 Method, apparatus and system for implementing fingerprint technology
CN101763437A (en) * 2010-02-10 2010-06-30 成都市华为赛门铁克科技有限公司 Method and device for realizing high-speed buffer storage
CN101820445A (en) * 2010-03-25 2010-09-01 南昌航空大学 Distribution method for two-dimensional tiles in object-based storage system
CN101866359A (en) * 2010-06-24 2010-10-20 北京航空航天大学 Small file storage and visit method in avicade file system
CN102004761A (en) * 2010-11-01 2011-04-06 青岛普加智能信息有限公司 Data storage method and system
CN102377972A (en) * 2010-08-23 2012-03-14 联想(北京)有限公司 Image processing equipment and method
CN104102646A (en) * 2013-04-07 2014-10-15 腾讯科技（深圳）有限公司 Method, device and system for processing data
CN104933067A (en) * 2014-03-19 2015-09-23 中国移动通信集团公司 Method and apparatus for operating file system and object storage system
CN105608193A (en) * 2015-12-23 2016-05-25 深圳市深信服电子科技有限公司 Data management method and apparatus for distributed file system
WO2016188063A1 (en) * 2015-05-25 2016-12-01 深圳市中兴微电子技术有限公司 Method and device for improving ram access efficiency, and computer storage medium
CN106250436A (en) * 2016-07-26 2016-12-21 东软集团股份有限公司 The method and device of management form data
CN107135264A (en) * 2017-05-12 2017-09-05 成都优孚达信息技术有限公司 Data-encoding scheme for embedded device
CN107851112A (en) * 2015-07-08 2018-03-27 云聚公司 For the system and method from camera secure transmission signal

Cited By (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101494658A (en) * 2008-01-24 2009-07-29 华为技术有限公司 Method, apparatus and system for implementing fingerprint technology
US8706746B2 (en) 2008-01-24 2014-04-22 Huawei Technologies Co., Ltd. Method, device, and system for realizing fingerprint technology
WO2011098017A1 (en) * 2010-02-10 2011-08-18 成都市华为赛门铁克科技有限公司 Method and device for realizing high-speed buffer storage
CN101763437B (en) * 2010-02-10 2013-03-27 华为数字技术（成都）有限公司 Method and device for realizing high-speed buffer storage
CN101763437A (en) * 2010-02-10 2010-06-30 成都市华为赛门铁克科技有限公司 Method and device for realizing high-speed buffer storage
CN101820445A (en) * 2010-03-25 2010-09-01 南昌航空大学 Distribution method for two-dimensional tiles in object-based storage system
CN101820445B (en) * 2010-03-25 2012-09-05 南昌航空大学 Distribution method for two-dimensional tiles in object-based storage system
CN101866359A (en) * 2010-06-24 2010-10-20 北京航空航天大学 Small file storage and visit method in avicade file system
CN102377972A (en) * 2010-08-23 2012-03-14 联想(北京)有限公司 Image processing equipment and method
CN102377972B (en) * 2010-08-23 2015-01-28 联想(北京)有限公司 Image processing equipment and method
CN102004761A (en) * 2010-11-01 2011-04-06 青岛普加智能信息有限公司 Data storage method and system
CN102004761B (en) * 2010-11-01 2012-12-26 青岛普加智能信息有限公司 Data storage method and system
CN104102646B (en) * 2013-04-07 2019-01-15 腾讯科技（深圳）有限公司 The method, apparatus and system of data processing
CN104102646A (en) * 2013-04-07 2014-10-15 腾讯科技（深圳）有限公司 Method, device and system for processing data
CN104933067A (en) * 2014-03-19 2015-09-23 中国移动通信集团公司 Method and apparatus for operating file system and object storage system
WO2016188063A1 (en) * 2015-05-25 2016-12-01 深圳市中兴微电子技术有限公司 Method and device for improving ram access efficiency, and computer storage medium
CN106293496A (en) * 2015-05-25 2017-01-04 深圳市中兴微电子技术有限公司 A kind of method and device improving RAM access efficiency
CN106293496B (en) * 2015-05-25 2019-05-31 深圳市中兴微电子技术有限公司 A kind of method and device improving RAM access efficiency
CN107851112A (en) * 2015-07-08 2018-03-27 云聚公司 For the system and method from camera secure transmission signal
CN105608193A (en) * 2015-12-23 2016-05-25 深圳市深信服电子科技有限公司 Data management method and apparatus for distributed file system
CN105608193B (en) * 2015-12-23 2019-03-26 深信服科技股份有限公司 The data managing method and device of distributed file system
CN106250436A (en) * 2016-07-26 2016-12-21 东软集团股份有限公司 The method and device of management form data
CN107135264A (en) * 2017-05-12 2017-09-05 成都优孚达信息技术有限公司 Data-encoding scheme for embedded device

Similar Documents

Publication Publication Date Title
CN1971562A (en) Distributing method of object faced to object storage system
CN100462979C (en) Distributed indesx file searching method, searching system and searching server
CN103067297B (en) A kind of dynamic load balancing method based on resource consumption prediction and device
CN102236581A (en) Mapping reduction method and system thereof for data center
CN102014042A (en) Web load balancing method, grid server and system
CN103810244A (en) Distributed data storage system expansion method based on data distribution
Yao et al. An energy-efficient and access latency optimized indexing scheme for wireless data broadcast
CN202033748U (en) Search engine performance test system
CN113900810A (en) Distributed graph processing method, system and storage medium
CN106709048A (en) High-performance computing-oriented distributed data organization method
WO2004063928A1 (en) Database load reducing system and load reducing program
CN103207920A (en) Parallel metadata acquisition system
CN109753244A (en) A kind of application method of Redis cluster
CN110245129A (en) Distributed global data deduplication method and device
CN107276914B (en) Self-service resource allocation scheduling method based on CMDB
CN102882960A (en) Method and device for transmitting resource files
US10375164B1 (en) Parallel storage system with burst buffer appliance for storage of partitioned key-value store across a plurality of storage tiers
CN102402544A (en) Attachment sharing method and device as well as business system
CN105138536A (en) Mobile social network data fragmentation method based on directed hypergraph
CN114090530A (en) Log summarizing and inquiring method and device under distributed architecture
CN109586970A (en) Resource allocation methods, apparatus and system
Liu et al. Sequential seeding for spreading in complex networks: Influence of the network topology
CN116405179A (en) Building networking data management method based on block chain slicing and DAG
Al-Yatama et al. Memory allocation algorithm for cloud services
Li et al. Blockchain transaction sharding algorithm based on account-weighted graph

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
C12 Rejection of a patent application after its publication
RJ01 Rejection of invention patent application after publication