CN109992796A

CN109992796A - A kind of MerCube machine translation management control system and method, computer program

Info

Publication number: CN109992796A
Application number: CN201910131256.9A
Authority: CN
Inventors: 米艳杰; 陶为民; 程国艮
Original assignee: Chinese Translation Language Through Polytron Technologies Inc
Current assignee: Chinese Translation Language Through Polytron Technologies Inc
Priority date: 2019-02-22
Filing date: 2019-02-22
Publication date: 2019-07-09
Anticipated expiration: 2039-02-22
Also published as: CN109992796B

Abstract

The invention belongs to the processing of natural language or switch technology field, a kind of MerCube machine translation management control system and method, computer program are disclosed；It include: application layer and podium level；Application layer includes: Integrated Management Module, updating and management module, monitoring management module；Podium level includes: device management module, Container Management module, machine turns over interface administration module, machine turns over Capacity Management module, copyright management module.MerCube provides more high bandwidth and more multilink, can promote the scalability of more GPU and more GPU/CPU system configurations.The present invention installs configuration using automatized script mode.Really realize one button installation, key upgrading downloading, start & shutdown through one key service, a key cutting system reaches light deployment, the convenient purpose of remote maintenance.The deployment of back-end services is using popular Docker container technique, so that system run all right, it is succinct convenient to manage.

Description

A kind of MerCube machine translation management control system and method, computer program

Technical field

The invention belongs to the processing of natural language or switch technology field more particularly to a kind of MerCube machine translation pipes Manage control system and method, computer program.

Background technique

Currently, the prior art commonly used in the trade be such that machine translation be it is a kind of using machine learning algorithm from bilingual Learn correlative connection between bilingual in panel data, recycles the rule of study that a kind of natural language is converted to other one The process of kind natural language.The development of the machine translation mothod development with the subjects such as computer technology, information theory, linguistics always Closely accompany.Machine translation lives through the phase of starting, phase of baffling, convalescence, the new period since generation twentieth century 30 or 40 years.Newly Period since nineteen ninety, experienced rule-based machine translation, the machine translation based on statistics and based on neural network Machine translation.Occupy dominant position at this stage is machine translation neural network based, compared to statistical machine translation it Using interpretative system end to end.Main thought based on neural network machine translation is using one " coding and decoding " Structure, coding are trained and are translated using a Recognition with Recurrent Neural Network respectively with decoder module.The sentence to be translated for one The sentence is converted to the vector of a fixed dimension by son, encoder first, and then using the vector as input, decoder can be given A string of term vector sequences out, it is eventually by the mode in dictionary lookup that the term vector of the output is Sequence Transformed for object language Word.

In existing mainstream machine translation, machine translation system whether based on statistics or neural network based Machine translation system is all that a GPU can only fill a natural language model, and structure and function are all relatively simple.

In conclusion problem of the existing technology is: in existing mainstream machine translation, whether based on statistics Machine translation system or machine translation system neural network based are all that a GPU can only fill a natural language model, Structure and function are all relatively simple.

Summary of the invention

In view of the problems of the existing technology, the present invention provides a kind of MerCube machine translation management control system and Method, computer program.

The invention is realized in this way a kind of MerCube machine translation management control system, the MerCube machine is turned over Translating management control system includes: application layer and podium level；

Application layer includes: Integrated Management Module, updating and management module, monitoring management module；

Podium level includes: device management module, Container Management module, machine turns over interface administration module, machine turns over Capacity Management mould Block, copyright management module.

Further, the Integrated Management Module includes:

With document process integrated unit and LangBox integrated unit and user's third party system integrated unit；

With document integrated unit: for responding the call request of " document processing system " in a manner of HTTP Post/Get, Return to translation result；

With LangBox integrated unit: in a manner of HTTP Post/Get, responding the call request of " LangageBox ", return Translation result；

With user's third party system integrated unit: being matchmaker in a manner of HTTP Post/Get, or with docx/execl file It is situated between, responds the call request of " user's third party system ", returns to translation result.

Further, the updating and management module includes:

It include: to upload administrative unit, download management unit, script administrative unit；

Upload administrative unit: for uploading file, data model, configuration file, program, the script etc. of upgrading；Wherein foot This document is to upload every time, it is necessary to include；After upload, the automatic perform script of MerCube completes the liter of back-end system Grade；

Download management unit: for any file in download log and MerCube system；

Script administrative unit: for being managed collectively all scripts for uploading, downloading, and the script that periodic cleaning is expired.

Further, the monitoring management module includes:

System Management Unit calls expense statistic unit, caller administrative unit, calls historical statistics unit, caller Rights management unit, equipment use monitoring unit, caller valuation administrative unit；

System Management Unit: it is managed for the system manager to MerCube system, comprising: title, password etc. Modification saves and sets system parameter；

Calling expense statistic unit: for interior for a period of time to user is called, the translation request of progress carries out expense system Meter, summarizes out in a tabular form；It can hour, day, the moon be that unit carries out language to number of words, the request number of words after, translation Statistics；

Caller administrative unit: for increasing to calling user, deleting, change, look into operation；

It calls historical statistics unit: for interior for a period of time, the translation request of progress to calling user, carrying out history and chase after Track, and with line chart, bar2d etc. is multi-form to be showed；

Caller rights management unit: controlling for that could can carry out translation request to calling user, including control The period of system, the language of control is to direction；

Equipment uses monitoring unit: for detecting to the equipment service condition in real time to GPU card, comprising: the past The busy occupancy of GPU in 1/4/8/24 hour；The number of request of language pair；Duration etc. is completed in average (translation) request；And with folding Line chart, bar2d etc. is multi-form to be showed；

Caller valuation administrative unit: when for that could can carry out translation request to calling user, carrying out charging control, Simultaneously for unit price (different charging unit prices can be arranged to different language pair in the future) to be arranged.

Further, the device management module includes:

GPU equipment distributes administrative unit, GPU develops environmental management unit；

GPU equipment distributes administrative unit: carrying out static allocation for GPU card, running what language model；Can also be When system operates normally, certain block GPU card is temporarily cut off, it is allocated；

GPU develops environmental management unit: for the exploitation environment to GPU card, drive module etc. upgraded, testing tube Reason.

Further, the Container Management module includes:

Machine turns over scheduling container unit, machine turns over engine container unit；

Machine turns over scheduling container unit: being used for start & shutdown through one key, manage and dispatch container, is equipped with Nginx (reversed generation in container Reason, static Web server), HAProxy (High Availabitity load balancing proxy server), Redis (accelerate to deposit by memory database Take), Mysql (database), uWSGI (Web server supports Python Flask frame)；

Machine turns over engine container unit: for start & shutdown through one key, management MT engine container；Be equipped in container: machine turns over Kernel program, module；The language model and configuration file that deep learning was trained.

Further, the machine turns over interface administration module and includes:

It calls rights management unit, call queue differentiated control unit, optimisation strategy administrative unit, flow control management single Member；

Call rights management unit: for allowing or forbidding user in certain time to calling user to carry out priority assignation It is interior to turn over system using machine；Meanwhile the use of language pair can be limited；In conjunction with current limiting measures, the use of user is also dynamically adjusted；

Call queue differentiated control unit: for carrying out priority (Priority) hierarchical management to calling user；Allow The user of high priority preferentially distributes resource, is responded；Only when the system free time, the user of low priority just can Shen It please arrive translated resources；

Optimisation strategy administrative unit: for Optimized Measures to be carried out with unified tactical management, Optimized Measures are combined It uses；

Flow control manages unit: for heavier in translation duties, when the request amount of user is larger, starting current limiting measures are come The equity for ensureing all users carries out limitation access to big request user, to guarantee that it is balanced that system is called.

Further, the Optimized Measures include:

(1) machine turns over the scheduling of resource, in conjunction with the management of the equipment of GPU video card, can both accomplish static allocation (card apparatus with Language model is fixed), it can also accomplish to dynamically distribute (card apparatus and language model are not fixed)；

(2) characteristic of redis memory database fast accessed is utilized, cache translations are as a result, so that next time directly uses；

(3) merge multiple small requests of user, it is disposable that backstage is submitted to translate, it returns the result and is submitted than simple sentence, speed is more Fastly.

Further, the Current limited Control includes:

Current-limit mode can be divided into static cost control, dynamic control, mixing control；

Static cost control: it is directed to certain user, sets fixed flow control threshold parameter (Filter or Criteria) in advance； Such as: request number of times < 10 time in 1 second；Request number of times < 100 in 1 minute；In 1 hour, the number of words of translation is no more than 1,000,000；When with After family reaches threshold value, just limits user and use a period of time；

Dynamic controls: translation interface, the time that backstage returns are called in detection every time, to judge the busy of backstage translation duties It is not busy；Such as: the response time of calling is necessarily less than some values (such as: in 20ms), is more than this value, then temporary current limliting；

Mixing control: while supporting dynamic control and static cost control；First dynamic, it is rear static；After dynamic examining passes through, then use Static method audit.

Further, the Current limited Control use includes:

Realize that flow control threshold parameter can be applied in combination using optimisation strategy administrative unit；Current limiting measures can be with simultaneously The rights management of user combines use, works well.

Further, the machine turns over Capacity Management module and includes:

Language is to administrative unit: cutting, customizes for that can turn over ability to the machine of MerCube system bearing；Which language To that can be exposed for users to use, which is temporarily disabled or hides speech；

Translation ability administrative unit: for being based on GPU equipment management, to the language of the GPU card load of MerCube system installation Say model carry out manual setting, to meet in business, in some period, some language to (such as: English en- > in zh) have it is larger The demand of translation.

Further, it the copyright management module: is protected using intellectual property of many kinds of measures to system；

(1) system acquisition MerCube machine key feature guarantees that engine can only operate on fixed hardware, and adaptation is solid Fixed GPU card；

(2) system is embedded in softdog the code of core translation algorithm and acquisition key feature algorithm, when operation, moves State load, anti-locking system are copied and are cracked.

Further, the MerCube machine translation management control method the following steps are included:

Step 1 merges multiple data packets according to request packet strategy is merged；

Step 2, design routing criterion are simultaneously scheduled data packet；

Step 3, processing data include middle text content, are placed in srcl_list list after subordinate sentence, and tgtl_list is arranged List is stored in the result for the correspondence sentence inquired in redis；

Step 4 loops through list srcl_list, searches from redis after value and MD5, the data of inquiry are deposited Storage is in tgtl_list；

The sentence that do not find is merged into one by step 5, continues to be sent to NMT translation；

Step 6 will be stored in redis after the sentence traversal after translation and MD5, and traverse and be filled into tgtl_list column In table, complete translation list is formed；

Translation list is spliced into character string and returned by step 7.

Further, in step 1, the merging request packet strategy includes:

{

“srcl”:”nen”,

“tgtl”:”nzh”,

" text ": [{ " APP_ID ": " adsadsd ", addr:11111, text: " content ", { }, { }]

}

Further, in step 2, the routing criterion includes:

Grade Priority dynamic is requested to be delimited, no series limits.Specific series is determined simultaneously by the application side of translation request Request；Rank is requested to be indicated with " single linked list ", request is indicated with " queue "；Wherein, N level highest, 0 rank are minimum；

(3) it is routed into rule:

When one request call reaches: first quickly searching language to (LangPair) chained list, after finding, in same rank Under, priority level (Priority) chained list is searched, after finding, request of newly arriving is come to the tail of the queue of all queued calls of the chained list； If language creates one, and be inserted in language to the end of chained list to not finding；If priority level is not found, One is then created, is arranged in a manner of descending, is inserted into position appropriate in priority level queue；After calling enters queue queue, It needs to count under this priority level, the Content-Length value of all-calls, and adds up, be then stored in priority level In node；

(4) rule is routed out:

1) language pair: there is no order of priority；All priority level chained lists of the language under are examined successively in order；

2) priority level: there is order of priority；First go out N grades of chained lists, after rolling, then N-1 grades out, and so on, to the last 0 Grade；With under level-one chained list, first Call_0, then Call_1 out out, and so on, to the last Call_n；

3) when request call reaches: any request call reaches, and checks the language where the calling to the excellent of node First rank obtains the Content-Length value of its preservation since N grades highest, and judges whether it has reached setting value Size: if had reached, begin to merge all call request packets, form one big packet, and issue；If no Reach the size of setting, then successively check N-1 grades down, and continues to build up the Content-Length value of its lower queue queue, In this way, to the last 0 grade；After packet after merging issues, need that simultaneously, the calling merged is emptied from queue queue；

4) timing length (the overtime duration of Waiting for Call) has arrived: all language are examined successively to section since correct Point checks treatment process, with the treatment process of [when a request call reaches] above；

5) check finally, if accumulation Content-Length value size less than the size of setting, will also remain Under queuing packet, merge, and issue；Meanwhile empty all language it is right/priority level it is all under queue queue.

Further, in step 2, the management and running include:

(1) request packet is received, different Request Priority queues, sweep spacing 5ms, packet length are stored according to priority parameters The data packet of >=2000 is transmitted directly to background process nonjoinder request；

(2) distribution mechanisms are triggered: since highest level queue, merging request packet since team's head, until reaching setting Packet length；

(3) merging is regular: trace interval then, only merges request of the same-language to direction, regardless of whether wrapping length Reach requirement, is all submitted to rear end and is handled.

The MerCube machine translation management control system is run another object of the present invention is to provide a kind of MerCube machine translation management control method, the MerCube machine translation management control method include:

Another object of the present invention is to provide a kind of calculating for realizing the MerCube machine translation management control method Machine program.

Another object of the present invention is to provide a kind of information for realizing the MerCube machine translation management control method Data processing terminal.

Another object of the present invention is to provide a kind of computer readable storage mediums, including instruction, when it is in computer When upper operation, so that computer executes the MerCube machine translation management control method.

In conclusion advantages of the present invention and good effect are as follows: merged multinomial leading machine in MerCube product Device translation system technology, comprising neural network machine translation technology, statistical machine translation technology, technical term translation technology and Translation memory library technology etc..Wherein neural network machine translation technology mainly uses the Machine Translation Model based on Attention Frame is machine translation structure of the novel encoder based on attention mechanism of one kind to decoder；Meet industrial application The data pre-processing and post-processing technology of standard greatly improve the accuracy of translation under the premise of guaranteeing processing speed.

MerCube using it is tall and handsome reach NVLink technology, more high bandwidth and more multilink are provided, can be promoted more GPU with it is more The scalability of GPU/CPU system configuration.The performance for promoting neural network translation system is improved using the technology, can be finally directed to The translation of whole chapter rank is accelerated.

The present invention installs configuration using automatized script mode.Really realize one button installation, key upgrading downloading, a key opens Withdraw business, a key cutting system reaches light deployment, the convenient purpose of remote maintenance.The deployment of back-end services is using prevalence Docker container technique, so that system run all right, it is succinct convenient to manage.

Configuration is installed using automatized script mode.Really realize one button installation, key upgrading downloading, start & shutdown through one key clothes Business, a key cutting system reach light deployment, the convenient purpose of remote maintenance.

The deployment of back-end services is using popular Docker container technique, so that system run all right, it is succinct convenient to manage.

For the optimum use for realizing translated resources, Mercube system uses a series of technologies, develops multiple functional modules Carry out scientific dispatch, system adjustment and optimization:

User is requested, multi-layer queuing, the request of preferential answering high priority are according to priority carried out；

Flow control, the unreasonable wink of restricted part user are carried out according to resource allocation conditions to the frequent requests of user Between largely request, prevent resource and call it is unbalance.

System also provides complete user management and surveillance and control measure, can carry out authority setting to user, be set using the time It is fixed, using language to setting, the service condition of user can be accomplished to show in real time, account of the history precisely counts.

General json interface, is easily integrated for convenience of MerCube system and other third party systems.For example it is turned over document System seamless interfacing is translated, rear module is shared, it is intensive to use resource.

The technological frame of MerCube system, typically with the development technique and deployment scheme of front and back end separation, so that View, data, structure mutually separate, so that data flow, service flow, page flow be made respectively to operate, safeguard, modify independently of one another, complete Beauteously realize Module Development Method

Integrated core translation engine mixes intelligent management (user management, equipment management, translation ability management), dynamic is supervised The functional modules such as control (operating status and service ability real time monitoring), charging (optional), so that machine, which turns over system really, becomes a It can easily issue, the independent machine of convenient management turns over system

Based on GPU equipment management, Mercube allows system manager to carry out dynamic adjustment to the language model of load, with Meet service dynamic management translation ability.

The protection of intellectual property is carried out using a variety of encryption measures.System provides soft encryption, and (core algorithm encryption storage is hard Disk, when execution, the decryption of dynamic importing memory), hardware encryption (core algorithm is implanted into softdog body) all multiple means are to protect It unites from the purpose for copying and cracking.

Detailed description of the invention

Fig. 1 is MerCube machine translation management control system structural schematic diagram provided in an embodiment of the present invention；

Fig. 2 is MerCube machine translation management control method flow chart provided in an embodiment of the present invention.

Fig. 3 is provided in an embodiment of the present invention to be routed into regular flow chart.

Fig. 4 is provided in an embodiment of the present invention to route out regular flow chart.

Fig. 5 is routing rule design diagram provided in an embodiment of the present invention.

Fig. 6 is that dynamic provided in an embodiment of the present invention changes GPU language to schematic diagram.

Fig. 7 is MerCube machine translation management control system functional frame composition provided in an embodiment of the present invention.

Fig. 8 is MerCube machine translation management control system Technical Architecture figure provided in an embodiment of the present invention.

In figure: 1, Integrated Management Module；2, updating and management module；3, monitoring management module；4, device management module；5, hold Device management module；6, machine turns over interface administration module；7, machine turns over Capacity Management module；8, copyright management module.

Specific embodiment

In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to Limit the present invention.

MerCube system, which provides, to fill the method for two natural language models on another block GPU card, and realizes Dynamic changes the function of language model.Furthermore translation maximum problem in cloud is the safety problem of data, and user is using cloud When translated product, cloud translation backstage can necessarily obtain the cleartext information of text to be translated, not utilize the protection of privacy of user. The appearance of MerCube system avoids problems；MerCube is the machine translation server of user-specific, using localization Deployment way, client can only access in a local network, ensure that user translates the safety of initial data from physical environment.

Application principle of the invention is explained in detail with reference to the accompanying drawing.

As shown in Figure 1, MerCube machine translation management control system provided in an embodiment of the present invention include: application layer and Podium level.

Application layer includes: Integrated Management Module 1, updating and management module 2, monitoring management module 3.

Podium level includes: device management module 4, Container Management module 5, machine turns over interface administration module 6, machine turns over Capacity Management Module 7, copyright management module 8.

Integrated Management Module 1 provided in an embodiment of the present invention includes:

Updating and management module 2 provided in an embodiment of the present invention includes:

Upload administrative unit: for uploading the file upgraded, data model, configuration file, program, script etc.；Wherein foot This document is to upload every time, it is necessary to include；After upload, the automatic perform script of MerCube completes the liter of back-end system Grade；

Download management unit: for any file in download log and MerCube system；

Monitoring management module 3 provided in an embodiment of the present invention includes:

Device management module 4 provided in an embodiment of the present invention includes:

Container Management module 5 provided in an embodiment of the present invention includes:

Machine provided in an embodiment of the present invention turns over interface administration module 6

Optimized Measures provided in an embodiment of the present invention include:

Current limited Control provided in an embodiment of the present invention includes:

Current limited Control use provided in an embodiment of the present invention includes:

Machine provided in an embodiment of the present invention turns over Capacity Management module 7

Copyright management module 8 provided in an embodiment of the present invention: it is protected using intellectual property of many kinds of measures to system；

As shown in Fig. 2, MerCube machine translation management control method provided in an embodiment of the present invention the following steps are included:

S101 merges multiple data packets according to request packet strategy is merged；

S102, design routing criterion are simultaneously scheduled data packet；

S103, processing data include middle text content, are placed in srcl_list list after subordinate sentence, and setting tgtl_list is arranged Table is stored in the result for the correspondence sentence inquired in redis；

S104 loops through list srcl_list, searches from redis after value and MD5, the data of inquiry are stored In tgtl_list；

The sentence that do not find is merged into one by S105, continues to be sent to NMT translation；

S106 will be stored in redis after the sentence traversal after translation and MD5, and traverse and be filled into tgtl_list list In, form complete translation list；

Translation list is spliced into character string and returned by S107.

In step S101, merging request packet strategy provided in an embodiment of the present invention includes:

{

“srcl”:”nen”,

“tgtl”:”nzh”,

" text ": [{ " APP_ID ": " adsadsd ", addr:11111, text: " content ", { }, { }]

}

As shown in Figures 3 to 5, in step S102, routing criterion provided in an embodiment of the present invention includes:

(5) it is routed into rule:

(6) rule is routed out:

In step S102, management and running provided in an embodiment of the present invention include:

Application principle of the invention is explained in detail combined with specific embodiments below.

One, merge multiple request packets:

1. merging request packet strategy

Merging data packet json format is as follows:

{

“srcl”:”nen”,

“tgtl”:”nzh”,

" text ": [{ " APP_ID ": " adsadsd ", addr:11111, text: " content ", { }, { }]

}

2, the merging data packet json format returned is as follows:

{

" text ": [{ addr:11111, text: " content " }, { }, { }],

" use_time ": " 0.1234 "

}

Two, criterion is routed

The design of translation request priority level is based on following several routing criterion:

1, it requests hierarchical, ordered

(1) request of all translations sorts by priority without exception, is up incremented by from 0,0 grade of most bottom；

(2) request of same priority is successive with the time sequencing reached, arrives first and is first lined up, is lined up after arriving afterwards.

Three, the design of criterion is routed

For the above routing criterion, design as follows:

1, request grade Priority dynamic delimited, and no series limits.Specific series is determined by the application side of translation request And it requests.

2, request rank is indicated with " single linked list ", and request is indicated with " queue ".Following figure indicates.

Wherein, N level highest, 0 rank are minimum.

It is routed into rule:

When one request call reaches:

Language is first quickly searched to (LangPair) chained list, after finding, under same rank, searches priority level (Priority) request of newly arriving after finding, is come the tail of the queue of all queued calls of the chained list by chained list；

If language creates one, and be inserted in language to the end of chained list to not finding；

If priority level is not found, one is created, is arranged in a manner of descending, is inserted into priority level queue and fit When position；

After calling enters queue queue, need to count under this priority level, the Content-Length value of all-calls, and It adds up, is then stored in priority level node.

Route out rule:

Language pair:

There is no order of priority.

All priority level chained lists of the language under are examined successively in order；

Priority level:

There is order of priority.

First go out N grades of chained lists, after rolling, then N-1 grades out, and so on, to the last 0 grade；

With under level-one chained list, first Call_0, then Call_1 out out, and so on, to the last Call_n

When one request call reaches:

Any request call reaches, and checks that the language where the calling to the priority level of node, is opened from highest N grades Begin, obtain the Content-Length value of its preservation, and judges whether it has reached the size of setting value:

If had reached, begin to merge all call request packets, forms one big packet, and issue.

If not reaching the size of setting, N-1 grades are successively checked down, and continue to build up its lower queue queue Content-Length value, in this way, to the last 0 grade.

After packet after merging issues, need that simultaneously, the calling merged is emptied from queue queue.

Timing length (the overtime duration of Waiting for Call) has arrived:

All language are examined successively since correct to node, treatment process are checked, with [a request call above When arrival] treatment process.

Note that check finally, if accumulation Content-Length value size less than setting size, also will Remaining queuing packet, merges, and issue.Meanwhile empty all language it is right/priority level it is all under queue queue.

Four, management and running:

1, request packet is received, different Request Priority queues, sweep spacing are stored according to priority parameters

5ms, packet length >=2000 data packet are transmitted directly to background process nonjoinder request.

2, distribution mechanisms are triggered: since highest level queue, merging request packet since team's head, until reaching setting packet It is long

3, merge rule: trace interval then, only merges request of the same-language to direction, regardless of whether packet length reaches To requirement, all it is submitted to rear end and is handled.

Five, GPU equipment detects

The detection of GPU equipment, including two aspect contents:

1, the detection of number of devices and type；

2, the detection of busy-idle condition；

(1) detection of GPU number of devices

It can detecte the quantity, title and UUID of the GPU card that system has been configured using nvidia-smi order.

And analog value is inserted into sys_settings table.

Insert into sys_settings(`key`,`value`,`memo`)

Values

(‘GPU 0’,‘GPU 0:Tesla P40’,‘GPU-dd57c03c-b961-3e93-638e- 1ed51f29552e’)

Insert into sys_settings(`key`,`value`,`memo`)

Values

(‘GPU 1’,‘GPU 1:Tesla P40’,‘GPU-d5f59f07-4c1d-923e-491a- 86d64af208fe’)

Detection method:

It needs outside container, executes detecting module probe_gpu_device to carry out.

Exploration policy:

After equipment installation, immobilize.Exploration policy positioning: it executes primary.

(2) detection of GPU equipment busy

It can detecte the busy snapshot of all GPU cards being configured in current system using nvidia-smi order.

And by snapshot results, after processing, it is inserted into snapshot flowing water table worksum_gpu.

Insert into worksum_gpu(`gpu_id`,`gpu`)values(‘0’,‘45’)

Insert into worksum_gpu(`gpu_id`,`gpu`)values(‘1’,‘23’)

Detection method:

It needs outside container, timing executes detecting module probe_gpu_idle to carry out.Fixed time interval can be by configuring File configures.It is recommended that detection in 1 second or 3 seconds is primary.

Exploration policy:

1) timing executes；

It 2) is to be not written into database when that is, value is 0 when the GPU free time convenient for counting and reducing data volume.

Six, increase gpu_langpair table newly for storing gpu card language to configuring condition:

1.1 table structures are as follows:

id gpu_name Langpair

1gpu 0EN-ZH

2gpu 1ZH-EN

1.2 call script execution gpu language to switching:

Script position:

1.3 newly-increased interface/uwsgi/admin/gpu_used, inquiry gpu utilization rate are shown on the page

Seven, systems increase interpretative function and start stop button, are placed on system parameter setting page

1.1 when pressing stop button, and system does not receive translation request, only allows management backstage page

When 1.2GPU card is arranged, it is necessary to which halt system translation (because to restart GPU configuration, prevents the request translated The data exception of return), it just can be carried out configuration

System function is restarted in 1.3 page increases, restarts background service program

Eight, increase GPU card newly and reset language model function

1.1 newly-increased front end pages " strategy setting ", the page shows that GPU can use card language model in current system, and works as The language model of preceding configuration

The language model reconfigured is sent to rear end by confirming button and carried out again by the optional GPU card model of 1.2 users Configure

The rear end 1.3Python increases interface newly to GP configuring U language model, requests nmt_worker_ by requests The flask server of 10260 ports carries out language model configuration in d0/nmt_worker_d1

It increases flask server in 1.4GPU docker newly, receives main program configuring request, perform script configuration language mould Type

Nine, configure haproxy

1.1 after configuration completes GPU language to setting, will reconfigure the direction of haproxy, when GPU language model is complete After portion's configuration successful, the script of configuration haproxy is executed, (problematic, view loses return value) (put in celery and execute)

1. the page (the IP address page) is arranged in system

Increase subnet mask setting, there is confirming button to be used to that IP and subnet mask is arranged

Increase translation start-stop button, after needing pop-up to confirm, page prompts user's interpretative function has stopped

Increase system reboot button, press rear pop-up " " restarting system, please refresh page is checked later ... " " it falls Timing 30 seconds (the above page, beautification beautification beautification)

2. all statistical graphs

Cake chart language statistics interface debugging on the left of 3.Dashboard

4. the preparation (python) of test data

5. softdog is tested

* increase user language and * is arranged to disabling to specified and language

Language is increased in subscriber frame newly to selection 6. adding, and user can choose multiple language pair

7. the modification page can also modify user language pair

It is arranged 8. user also can be set in disabling setting for the disabling of language pair

The rear end 9.Python is for different user's different languages to storage blocking information

10. the caller center page will show the available language of user to (graceful display), user disables mark and (need to set Meter is placed on reasonable position)

It should be noted that embodiments of the present invention can be realized by the combination of hardware, software or software and hardware. Hardware components can use special logic to realize；Software section can store in memory, by instruction execution system appropriate System, such as microprocessor or special designs hardware execute.It will be understood by those skilled in the art that above-mentioned equipment Computer executable instructions can be used and/or be included in the processor control code with method and realize, such as in such as magnetic Disk, the mounting medium of CD or DVD-ROM, such as read-only memory (firmware) programmable memory or such as optics or electricity Such code is provided in the data medium of subsignal carrier.Equipment and its module of the invention can be by such as ultra-large The semiconductor or such as field programmable gate array of integrated circuit or gate array, logic chip, transistor etc. can be compiled The hardware circuit realization of the programmable hardware device of journey logical device etc., can also be soft with being executed by various types of processors Part is realized, can also be realized by the combination such as firmware of above-mentioned hardware circuit and software.

The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.

Claims

1. a kind of MerCube machine translation management control system, which is characterized in that the MerCube machine translation management control System includes: application layer and podium level；

Podium level includes: device management module, Container Management module, machine turns over interface administration module, machine turns over Capacity Management module, version Weigh management module.

2. MerCube machine translation management control system as described in claim 1, which is characterized in that the integrated management mould Block includes:

With document integrated unit: for responding the call request of " document processing system " in a manner of HTTP Post/Get, returning Translation result；

With LangBox integrated unit: in a manner of HTTP Post/Get, responding the call request of " LangageBox ", return to translation As a result；

With user's third party system integrated unit: in a manner of HTTP Post/Get, or using docx/execl file as medium, The call request of " user's third party system " is responded, translation result is returned.

3. MerCube machine translation management control system as described in claim 1, which is characterized in that the updating and management mould Block includes:

Upload administrative unit: for uploading file, data model, configuration file, program, the script etc. of upgrading；Wherein script text Part is to upload every time, it is necessary to include；After upload, the automatic perform script of MerCube completes the upgrading of back-end system；

Download management unit: for any file in download log and MerCube system；

4. MerCube machine translation management control system as described in claim 1, which is characterized in that the monitoring management mould Block includes:

System Management Unit: it is managed for the system manager to MerCube system, comprising: title, password etc. are repaired Change, save and set system parameter；

Calling expense statistic unit: for interior for a period of time to user is called, the translation request of progress carries out expense statistics, with Form summarizes out；By hour, day, as unit of the moon carry out language to, translation after number of words, request number of words statistics；

Call historical statistics unit: for interior for a period of time to user is called, the translation request of progress carries out historical tracking；

Caller rights management unit: it is controlled for translation request could can be carried out to calling user, including control Period, the language of control is to direction；

Equipment uses monitoring unit: for detecting to the equipment service condition in real time to GPU card, comprising: the past 1/4/ The busy occupancy of GPU in 8/24 hour；The number of request of language pair；Average request completes duration；

Caller valuation administrative unit: when for translation request could can be carried out to calling user, charging control is carried out, simultaneously For unit price to be arranged.

5. MerCube machine translation management control system as described in claim 1, which is characterized in that the equipment management mould Block includes:

GPU equipment distributes administrative unit: carrying out static allocation for GPU card, running what language model；Can also system just Often when operation, certain block GPU card is temporarily cut off, is allocated；

GPU develops environmental management unit: for the exploitation environment to GPU card, drive module upgraded, test and management；

The Container Management module includes:

Machine turns over scheduling container unit: it is used for start & shutdown through one key, manage and dispatch container, Nginx, HAProxy are installed in container, Redis, Mysql, uWSGI；

Machine turns over engine container unit: for start & shutdown through one key, management MT engine container；Be equipped in container: machine turns over core Program, module；The language model and configuration file that deep learning was trained；

The machine turns over interface administration module

Call rights management unit: for allowing or user being forbidden to make within certain time to calling user to carry out priority assignation System is turned over machine；Meanwhile the use of language pair can be limited；In conjunction with current limiting measures, the use of user is also dynamically adjusted；

Call queue differentiated control unit: for carrying out priority level management to calling user；Allow the user of high priority Preferential distribution resource, is responded；In the system free time, the user of low priority can just apply to translated resources；

Optimisation strategy administrative unit: for Optimized Measures to be carried out with unified tactical management, use is combined to Optimized Measures；

Flow control manages unit: for heavier in translation duties, when the request amount of user is larger, starting current limiting measures to ensure The equity of all users carries out limitation access to big request user；

The Optimized Measures include:

(1) machine turns over the scheduling of resource, in conjunction with the management of the equipment of GPU video card；

(2) characteristic of redis memory database fast accessed, cache translations result are utilized；

(3) merge multiple small requests of user, it is disposable that backstage is submitted to translate, it returns the result and is submitted than simple sentence；

The Current limited Control includes: that current-limit mode is divided into static cost control, dynamic control, mixing control；

Static cost control: it is directed to certain user, sets fixed flow control threshold parameter in advance；After user reaches threshold value, just limit User uses a period of time；

Dynamic controls: translation interface is called in detection every time, the time that backstage returns, judges the busy of backstage translation duties；It calls Response time be necessarily less than some value, be more than this value, then temporary current limliting；

Mixing control: while supporting dynamic control and static cost control；First dynamic, it is rear static；After dynamic examining passes through, then with static state Method audit；

The Current limited Control use includes:

Realize that flow control threshold parameter can be applied in combination using optimisation strategy administrative unit；Simultaneously current limiting measures can and user Rights management combine use；

The machine turns over Capacity Management module

Language is to administrative unit: cutting, customizes for that can turn over ability to the machine of MerCube system bearing；Which language pair It can be exposed for users to use, which is temporarily disabled or hides；

Translation ability administrative unit: for being based on GPU equipment management, to the language mould of the GPU card load of MerCube system installation Type carries out manual setting, to meet in business, in some period, and demand of some language to there is larger translation；

The copyright management module includes: to be protected using intellectual property of many kinds of measures to system；

(1) system acquisition MerCube machine key feature guarantees that engine can only operate on fixed hardware, is adapted to fixed GPU card；

(2) system is embedded in softdog the code of core translation algorithm and acquisition key feature algorithm, when operation, dynamic plus It carries, anti-locking system is copied and is cracked.

6. a kind of MerCube machine translation management control of MerCube machine translation management control system described in operation claim 1 Method processed, which is characterized in that the MerCube machine translation management control method the following steps are included:

Step 2, design routing criterion are simultaneously scheduled data packet；

Step 3, processing data include middle text content, are placed in srcl_list list after subordinate sentence, and tgtl_list list is arranged It is stored in the result for the correspondence sentence inquired in redis；

Step 4 loops through list srcl_list, searches from redis after value and MD5, the data of inquiry are stored in In tgtl_list；

Step 6 will be stored in redis after the sentence traversal after translation and MD5, and traverse and be filled into tgtl_list list In, form complete translation list；

Translation list is spliced into character string and returned by step 7.

7. MerCube machine translation management control method as claimed in claim 6, which is characterized in that in the step 2 Routing criterion includes:

Grade Priority dynamic is requested to be delimited, no series limits；Specific series is determined and is requested by the application side of translation request； Rank is requested to be indicated with " single linked list ", request is indicated with " queue "；Wherein, N level highest, 0 rank are minimum；

(1) it is routed into rule: when a request call reaches: first quickly searching language to LangPair) chained list, after finding, Under same rank, priority level Priority chained list is searched, after finding, the request that will newly arrive comes all queued calls of the chained list Tail of the queue；If language creates one, and be inserted in language to the end of chained list to not finding；If priority level does not have It finds, then creates one, arranged in a manner of descending, be inserted into position appropriate in priority level queue；Calling, which enters, is lined up team It after column, needs to count under this priority level, the Content-Length value of all-calls, and adds up, be then stored in excellent In first level node；

(2) rule is routed out:

2) priority level: there is order of priority；First go out N grades of chained lists, after rolling, then N-1 grades out, and so on, to the last 0 grade； With under level-one chained list, first Call_0, then Call_1 out out, and so on, to the last Call_n；

3) when request call reaches: any request call reaches, and checks the language where the calling to the priority of node Not, since N grades highest, the Content-Length value of its preservation is obtained, and judges whether it has reached the big of setting value It is small: if had reached, to begin to merge all call request packets, form one big packet, and issue；If do not reached The size of setting then successively checks N-1 grades down, and continues to build up the Content-Length value of its lower queue queue, in this way, To the last 0 grade；After packet after merging issues, need that simultaneously, the calling merged is emptied from queue queue；

4) timing length has arrived: all language being examined successively since correct to node, treatment process are checked, with above The treatment process of [when a request call reaches]；

5) check finally, if accumulation Content-Length value size less than the size of setting, also will be remaining It is lined up packet, is merged, and issue；Meanwhile empty all language it is right/priority level it is all under queue queue；

In the step 2, the management and running include:

(1) request packet is received, is stored in different Request Priority queues according to priority parameters, sweep spacing 5ms, packet length >= 2000 data packet is transmitted directly to background process nonjoinder request；

(3) merge rule: trace interval then, only merges request of the same-language to direction, regardless of whether packet length reaches It is handled it is required that being submitted to rear end.

8. a kind of computer journey for realizing MerCube machine translation management control method described in claim 6~7 any one Sequence.

9. a kind of realize at the information data of MerCube machine translation management control method described in claim 6~7 any one Manage terminal.

10. a kind of computer readable storage medium, including instruction, when run on a computer, so that computer executes such as MerCube machine translation management control method described in claim 6~7 any one.