From 9572e11d73ede8040d6495f8e588b61fb9ea20ed Mon Sep 17 00:00:00 2001 From: Helin Wang Date: Mon, 1 May 2017 18:27:28 -0700 Subject: [PATCH] Design doc: master process --- doc/design/cluster_train/data_dispatch.md | 6 +- doc/design/cluster_train/master_process.md | 89 +++++++++++++++++++ doc/design/cluster_train/src/dataset.graffle | Bin 0 -> 2770 bytes doc/design/cluster_train/src/dataset.png | Bin 0 -> 10845 bytes 4 files changed, 92 insertions(+), 3 deletions(-) create mode 100644 doc/design/cluster_train/master_process.md create mode 100644 doc/design/cluster_train/src/dataset.graffle create mode 100644 doc/design/cluster_train/src/dataset.png diff --git a/doc/design/cluster_train/data_dispatch.md b/doc/design/cluster_train/data_dispatch.md index a3eb4e28d..f60c3b843 100644 --- a/doc/design/cluster_train/data_dispatch.md +++ b/doc/design/cluster_train/data_dispatch.md @@ -21,7 +21,7 @@ ### 文件预处理 -在数据集可以被训练之前，文件需要预先被转换成PaddlePaddle集群内部的存储格式（SSTable）。我们提供两个转换方式： +在数据集可以被训练之前，文件需要预先被转换成PaddlePaddle集群内部的存储格式（RecordIO）。我们提供两个转换方式： - 提供给用户本地转换的库，用户可以编写程序完成转换。 - 用户可以上传自己的数据集，在集群运行MapReduce job完成转换。 @@ -92,11 +92,11 @@ random_images-00099-of-00099 #### 进行训练 -PaddlePaddle提供专用的[data reader creator](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/reader/README.md#python-data-reader-design-doc)，生成给定SSTable文件对应的data reader。**无论在本地还是在云端，reader的使用方式都是一致的**： +PaddlePaddle提供专用的[data reader creator](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/reader/README.md#python-data-reader-design-doc)，生成给定RecordIO文件对应的data reader。**无论在本地还是在云端，reader的使用方式都是一致的**： ```python # ... -reader = paddle.reader.creator.SSTable("/home/random_images-*-of-*") +reader = paddle.reader.creator.RecordIO("/home/random_images-*-of-*") batch_reader = paddle.batch(paddle.dataset.mnist.train(), 128) trainer.train(batch_reader, ...) ``` diff --git a/doc/design/cluster_train/master_process.md b/doc/design/cluster_train/master_process.md new file mode 100644 index 000000000..949811b4f --- /dev/null +++ b/doc/design/cluster_train/master_process.md @@ -0,0 +1,89 @@ +# Design Doc: Master Process + +For an overview of master process' role, please refer to [distributed training design doc](./README.md). In this design doc we will discuss the master process in more details. The master will be implemented in [golang](https://golang.org/). + +## Dataset + + + +A dataset is represented by a list of files in *RecordIO* format on the distributed filesystem, each RecordIO file consists of multiple *blocks*, and each block has multiple data instances. + +## Task Queue + +As mentioned in [distributed training design doc](./README.md), a *task* is a data shard that the master process assigns to the trainer process to train on. A task consists of one or multiple *blocks* from one or multiple files. The master process maintains *task queues* to track the training progress. + +### Task Queue Creation + +1. Each trainer will make an RPC call (using [golang rpc](https://golang.org/pkg/net/rpc/)) to the master process, telling it the RecordIO files representing the dataset specified by the user. Since every trainer will tell the master process the same dataset, only the first RPC call will be honored. + + The RPC interface is: + ```go + func (m *RPCServer) ReportDataset(Paths []string, dummy *int) error { + } + ``` +1. The master process will scan through each RecordIO file to generate the *block index* and know how many blocks does each file have. A block can be referenced by the file path and the index of the block within the file. The block index is in memory data structure that enables fast access to each block, and the index of the block with the file is an integer start from 0, representing the n-th block within the file. + + The definition of the block is: + ```go + type Block struct { + Idx int // index of the block within the file + Path string + Index recordio.Index // block index + } + ``` +1. Blocks are grouped into tasks, and tasks are filled into the todo queue. The pending queue and the done queue are initialized with no element. + + The definition of the task is: + ```go + type Task struct { + Index int + Blocks []Block + } + ``` + + The elements in the tasks queues is of type `TaskEntry`, containing a timeout counter (described in [task retry logic](#task-retry-logic)), and a task: + ```go + type TaskEntry struct { + NumTimeout int + Task Task + } + ``` + + The definition of task queues is: + ```go + type TaskQueues struct { + Todo []TaskEntry + Pending map[int]TaskEntry // map from task index to task entry + Done []TaskEntry + } + ``` + +### Task Queue Persistence + +The task queues need to be persisted on [etcd](https://github.com/coreos/etcd) for fault recovery. Since the task queues only change once a task is completed or timed out, which is not very frequent, we can afford to synchronize with etcd every time the task queues change. + +We will serialize the task queues data structure with [gob encoding](https://golang.org/pkg/encoding/gob/), compress with gzip, and save into etcd synchronously under key `/task_queues`. + +### Task Dispatch + +The trainer will make an RPC call to master to get a new task when: + +- the trainer first started, or +- the trainer finishes a task. + +The RPC interface is: +```go +func (m *RPCServer) GetTask(finished *Task, result *Task) error { +} +``` +Argument `finished` will be `nil` when the trainer is just started. + +During the RPC call the master will do the following: + +- Make a copy of the task queues, and update the copy reflecting the finished tasks and the new pending tasks. +- Synchronize the copy of task queues with etcd using a transaction conditioned on holding the master lock. +- Replace the task queues with the copy and report to the trainer with the new tasks if succeeded, or discard the copy and report the error to the trainer if failed. + +### Task Retry Logic + +When a task is dispatched to the trainer, the master will schedule a function for execution after the timeout duration (based on the moving average of task completion time). If the task entry in still in the pending queue, its timeout counter will increase by one, and the task will be moved to todo queue. If the timeout counter is above the threshold, the master will log the error and discard the task. diff --git a/doc/design/cluster_train/src/dataset.graffle b/doc/design/cluster_train/src/dataset.graffle new file mode 100644 index 0000000000000000000000000000000000000000..c10a423ed16a23229a9ee33d11bfc82bb59646c8 GIT binary patch literal 2770 zcmV;@3N7^?iwFP!000030PS3BbK1xf{T%-at-r1fk^lo)$6F)bj*ZRB1d6I#gESx+ zJ%}WXNy`8I_UHk6fyvG0+NqVRj2?4(rl&Q1x;4gq`S-J?GEc;`6hr@X%4OJ7hUlUp zDf;}=>7ZXfFP;8GmOJdra8|dsD1TvB>ox(6~?;&;IuB3?e z)yap;1=(C7(N>)JZR4sbLDbAfVj;1J-ZsL zYy(M;E6a|l%N1h!3Rqmw@pTEtUdoEN74i(>!6?G=+qOD=|9;M~Jb%WVm-yn#8S_2I zUR<0pe7?lJyo4q}1uH7H;(hdFyd=BPBQmY(PQX{456;4c_3VvGc1T}$w-LHD7Q5C! zv)f)Z@p`A%PP6u1P#v)I)+GMf?!>jFMaG7q9kEO*SbxISovF-~-2(BA%APT7(0T!h zp7M>v@a`j^Paq~_s4GoJsxsH(iy^L*gYGS{2wiOPCc)}e?&31*ocK_W%QSu0<{Sf!t4?))rS#66B zyKiJhE@Y;tY7h~IY!~ouIVGW7*fb3BmN=U!l*2!z?V1lt&wonV{3&S*e3L|b@QiPn z^J(Ns1UW0=%IQ(YlCgfxuu+EGPmw==w-o0?T+TVZ zz)wU&G%#(-vw8QzT20|PUt~|dPiBU0+f#KiV<$6Rfuks6(;_y@0Mk4X8^U#)tjJ$4 zfEDQq0`OlGlSq7_FqxZ#==5Z&t_U0fHCbX)n%K{U3l=^(j&}|hC3@ibymKhA^pMLr zkX+e;2%Y_CkLPQHrthONwHo9Htv0#mXlC(bOa9M4R| z8OsQ&AueK`8!&n3{C{ByzF2qzmdK|pk+MX}62A{i@VPf&i9*T}DNCd*@%~wYEuJx4 znR6(jP|6hO75o2UiHmp45+yeLlm8)>I8NI9Dd~R)ON4ev!s~~PE}xWFcD(W`snBhg zL!7cR=Di9Fjc&LkdU5nA1p3%*eiqS zdiSZ6&k}F4EyI;lmPlD5Wr^R1CEjFPhD})_Wr>s}-akv^a_&~l8*R*R>BdaDF_Ui0 z{L43HxZ*qCnBgzVZ})wf9G50jX)={2Q|~^R@>$|dHfD0^D_6=ADNFoDEb%5AGwJqi z$`Y@!M7LkBMl`}U_Adlvf0LTm>S)NZkLn)K*9Mynd)m8|6ow$2Fv>*B%6Q zhCLa^t=0`R6B~F-Odw5nJeY__8(4 zwGbU_w7!Eu#u`G6h6h;~y`d@#cLPq4MXj#G?;6+JEEztl{yPyOuqO*pR_gsx^ig$f#TGutFcrD|OVu zJ%6BUJQNl%VXeH9TJB8zYu|Lj#X#%9?FZ{yTWd?g93dpE;iH=MAlbznjXQ#cnj?B- zQDcEdYe7b}y#I{gHixLHsIJwnEu^g*Qv^rZz>VEn3oMQIA7f|(>)vLao9ufEJNz#V zJM0_o^Oe$lzI31OAK&MDmw~F=zz`H0Uq1kiBU^B{ioqcad{?P{lfi%<)K1*q&nAmlbc<;IjRu-;6z*B+#~?2WHm zo%--q#yU$KRrkf2`Xx=ku-;6vx9I)ctnjF>ch2fy5;~;Rnc(N(bknI z8q;7^X~@RtQyQrL>5XIl9etzk$c^J@U<`!Cd=3I(fzLQwHU;`J2Dc(%@zCb;{*bs| z?XcE~u`v4o$uYu?1^*aX@1yFFmBV%lyF^rnC_(RkjtS%(pyorw9A#ujAV|xH_?TiG zfgKx5wGrgvBLBepj4TwwRO|=@`S=hKQwDZN*h9lf{Obt#Bz%Z{ecE*=oxr{ONcY5l z0XamuL9TTKqJ;Y@)4s~HL6%M8*py+9WY{X05oY$Xr})4@g5Bemfr160S`iIXky`r9 zNG|$7TO-u~yyTVJrK%3{7C#y$g@@De2ovKtgR&Ve&YRN&TpTq&3S}OJiX_?l^6Nm1 z3C%iYB;WL*|6>goi)gGAiC8!A%15MVQGyU_2hb0SrA*_c4@R|K6yLnZb7IAc7Vxo} z1uYGL6$p_wF^?z_@tKfi-v}3Am|@i-Ss&q4OR-)pe}KS2!QDfFdE1ePp~gAT+{RkM zGjqmg5?0kcTn0VGHdn#k9u%@pYTf5LFpjaZ_d5^L6)jL4V^2hOprT@m#;u}P19c;O zM=M1<&B68vkcd-%BLZhLYy_`5YS#3*qQ`ZS3_GTbPHS@90Q!dLb6buLz&M7?! z@LU(x5m@TD`yH*_6Aarnv|DTjis=Nu@8NenQw?l~iNQP1vtgRwv6eeR(=e59hK^Ox zThSv6H9nANNR-qQ20EDjA{jEn0yiCAtB_v8HDrq1Se`Z?R0Qo6NG<#P_17^ZuBbJv zq25=()Sp$z!Var{LuU5Rn1`~^q5kjTHOLH>CQ9nwa@rGm|9-*oXH2oozq~RET?bu4 z3wDu&#rZoth%Qb1e&v(s8_QqOeVBTk2XUtp>co4pfG)w(ovvCYY=F?cJ41%)J13LR0)&^H7n3+0X938xm&ZiTtOJ zr$<@9vwoM%L;SV11bfDu=iI;6Y)iYv4G}Cgx97c#LLM-@-&%=$5d2Eyeh>MO(0kIY YAREP)V|R4QE?%Ad7qtPw^$k)00GL^19RL6T literal 0 HcmV?d00001 diff --git a/doc/design/cluster_train/src/dataset.png b/doc/design/cluster_train/src/dataset.png new file mode 100644 index 0000000000000000000000000000000000000000..2fb7f1cce3b6dd21489392557826e95a9f207c34 GIT binary patch literal 10845 zcmeI21yt03_vZ%$1q6hlq`QS7MUa&4Zb@mRQ=~&e5u`(g2FU>dK@boTk?xR?knWad zKZC!0p8xLIXV3FL=f7w7?D=_)Iy2l)-S7R}`?~KTQcXn`2a5~~0)gPj%SmZKAgGk! z^#KMd__t#8#t`@q#Z5!@5v07AViml>be7Y1gFtQ*ApfF3(%#+$1|zmwdhUA4Nd~NwZeD=?E z{@AP8IJr9kFSy!1k#}^rbOnRm&5*kj<@sy$zqa_F=TddGwFIZW9?kvN=>PcHU-rT< zd!#X$XXcR$fX%%Nu1g<5>mnkCS#c(la5?VEpxQ z#_|kq&#(xQ+*npiQaUq**tEM6q;V1a7Q-Bpok@7>G)Zbnl~QjoE4SnH9&McUZa;MM zA9Ct@ZnD()d(+=z=JHd(Cy&#E)uCly_%hrnbN{XXb}d{i03I8=!y$?n2tlKP1f^ig zsTpfB#u59Omq%mKKu|EmX;#qk*jIyl11}{ZsCa>axoG`FvMw5=Nb}DG;<~8Ys1W4f zSr!!cHubk!H<4p@(E1ZUH$WiB86GG=j9dtr6D7cK3}F;^#vM)w1VbEx`WhNh%I?&B z@!^OjUG+7EVAJo$?Zd6g%D$%PNa8U3s`xz==@b-< zPtHxy7cvGscEj(gtozCzf7B>+oN2lk%Dk+$8OR~!wetNHE%5C%Wl8t1_L#AE>c3`N zLSB-bW?4}O)}qL!vm`>? zR|i?YXIFpW2sj`JoieS5Mpkh`Szhz_m&HRVvY zmNl4x>1rTX7Fwn%cA*tMUGJeev9NB)pQoC6Ip6kDLG|qeb%FOPZ?0n&1%cmBbM@s| zK?)}!XQ4O4(B4GA=_en0wfqJm;9@1Ey}eba`>s8E}4FiOO4L)x%@!r&#NM zbv9$&lU8HY;0eqwH>&!&7mGJ##ZCv7GXorPn=dw783%nj`GLdf=clL)hV4Qr={di)<-Sa@tLCc_%M;AV1u2EBjws?H zzCN{cdUIx?ZiG?8hc6rxzw=QGRWf+m2s@mn>kq{)cVnV)Mx!jii~9D^@rf!@s$7|Fnah$H__gK^N5P#To zP{wuE`205^|H<5)0Po*hMm@rC;b%F%$`5d3g%bU0 zbKSOc`36Wy5L7C+7UM$u)_$Ll={t+ zI~DI}&E5$(4#B75aF)N{nZ#jVtYLn3v>Da9u0~5>8;K94v};c`)*EPGE$|-!&f8;D zkD8Y=Qo39nD(p&N%_SW?+irGpJ$`Y0IaKlc`!lj`%F`Z{g%J8N+(>I#$(N(&N|Ti1 zRhlr4u^XLCNV8iJkzKSOP};v6KbmbNn09ydf@n4Vf-Rq@WX#yYK@+!Y84sji2%4~> zI9sfb(e6`(&D}rQi48R_k>s1Mv37bL?D~`AG)Vot{r-SA^4M_tz;f?i20jdA@3jsyw z>P&Ht*Nu{xqG2MVtKZF6)5T#@X|%B^DJeNkvo~44Earak4vu-XXsoHZcC&!7G~}Z- zz8&&O;7OsbDBzW;WYU5Y*U5si6vM%ftZEo^pe#k`6!swj!mg7|> z(L_&Iavy(m{c)Be^4xQg&vEL3V$!`M-$C0H$c;y2_qVzlmh#Hwq}PUv;?Mo5{g2J| zlw1e7Piy+^F>%Z2?l## z_6#N$&GB^b%=o4EO6Tscc7$;w*=Rz(V)E;|+&Q2!Q2XvkY)yR($0q#(3l61KcoP15 zYx<$`Pie!UxKHzF5yvPUMW`MGw&t|pjD)CBsnO`_&IP^v$gH#Q)Umw z)6s>Vblor_b4*SADwyD@56h|->Ru>7m9R~MDn#~4V@S0a8Hla{wLSIJemZYqa-#wz z(r8L##@zQkJUQR*t6^m^=VF56{4#c7%F;&gTKBHxCm|?k`?^zkL9wJ*wsLOJRSZO# z8`<+Z-$Kaf@=1&q49=V{2^bjbDoM`}FsqbmSSWzlX4^d@f5-_b?_&NwD>aHq3ZEV4- z1?@r%-uEJf)#k}*L@zFb?-Nl=KA%rWS(0|1_dgF=0RY)bL z%YsKdV3jqIx<_GHZP_JLbY!R#Ym|+=4;F9QrTlb!cDysUj%cP*re`I%g^Kt}0Om0L zK{!n(?PzK@TCvl)#9n}o^^S&(mB-D1qcrlRsBMkc0lNb5oqF(|ccm_Y$?gNfv|HEJ zqKGN5Jdcc$NJqk-XgRt51$t9coSQ6#zDe{PY-C(|8u zHKM2F(@lQQ0TmtXrSe#^J+5<3<}^}BWXnGX_b!@R%ztCN@?>wZisYrJ$kXu(Cgn5$ z-E^mlETjRb=>ks1bR zzM^6>r-$Nq7L&j$_ zTAHdeQd-4S`J^3}HR)|0NO)$24-UvsZlV?5gHF|5oa`@;l^G_3D;~$F_+sFb_ihex zd%1CA@m$uk{iXI8YTc>DpN-Gi3QG9eJ_cO*1BfRVaP$S25$u)RtD5bj>tbh~d^}a;$yMg8#-RA?*6GZv@7jn z&}%^xLG&E0kHMbxP+35`nW9N~zrBl~A|W&%ShHE~&1hBIm~PNTIR`DH_B%C+Lc7sY z)-F3@LvT>wa5?F;`wn*HaZIeT`*EPr(NA0nXy{Gll?@)QFq34D4=u#tqN%fZZMqso zTyPJiKk&WNOj9HQvjT+ShG$b+n9V(|oSbvvXM5l=OQ#MJ`3cw!e}`gkqW04}e0xRO zvp8`(D_fR--Q>A}9S9-41-7Q@`Yim*w_mjcv#^Ixrv_!V-Teu~fP=*HPd;0IC)=ScMSq60~OeIS0 zVyc3R>D)ol5{u_+#? zF%p?nj>!Y#yLfNkt7Lgd-(H6jIV@EU4Ik503OdgQCKj1?zoC~+NaZvVa!?xHi||lL zW`C4RY0gv#5)2fMODXhSg5;i_DG~jg3+j6W@z)LByZm~@y7zRe2ugPMuvnv*xT8u~ zhqf@gl3|%#X2Fp^RuZT}YkCW2qpB4qzm<~7d8a~d1Z$3BCY7&5?)B?ZJ3vN`sFeAu z8K5$nXLZ$UHQu}Co17!nPmVTz7-%^Cc%L6-3oE)wvWY%;R!{b-`nFsNg--Xk;arj} z!rqarTq7AXezz5TR3F=p2f;^p_O5~fm%cqJK&>!yf`7mT9WFmB>*~gjDxVf<*qXgl zN|FG7#>lU*oZnGd-)}8o?&W7Z^a59GY>-GZ=8IEQchPv6VeNHjfh3TOjbD1)X>H<< ztq_gtv!?ZK^iz=2atUoNxWI!%n<$!1!L|T9xFag zb@5(d!?ZY+^!uA2jl>bO?bMXD`hS$S?*uXo%d){B-MKJ&SjCO;M(b7jQTR>d*T>6 z*P2kdBvg{nWLRO?G=F>9UDl40!Y@3H>}zJBKzd`W(7KJR-dUh}2a>@Zk$sJGWFULo8RB;O31*4SSax$)9~b7%Y1%GBzi!>tpeC3w8VMH zm*O;@fQdG+i%`I>A!v_)8QyJ;KQliB2;uo+7-M`Qq~dMIfJ27pEJ zP?iIz?{RZ;Yl*O`LL>TA@6Tbv`7X!Eh;FHV#%Yq7iof-`A!xsz&QV@o9-b$g+eCxc zN1fb8@M*)%Wgl&d#>VlpP)ag~$|Q0_$1$Ke7E2#6N3LWHD9`(cP&E$H?^j^gmsO-3 z*DkdLzD_Qa{69XV{Cq3#)C^rBJ4^~&Q_;X&jn~-fD?&f#-BXaN1+TQB%mvt0ZJ?I zn(J~eQbPs#$X2^Y$R7I8b0gNgL%uT^M(r6Ti9i;mPP&%p$xN{Te}Ha)gkS(J@AiP^ z(9UH8^wMOwX#!6E=f^H>MxaO*n%666z6I>!1(+3ZWcVFwxTs&fi)=|8AXB9@KKWg8 z97!D}OebipD}{)UJMc!0{TQw$I!6+ad8Iqmfr~&pso=q+3jYL9YUmA+8T--&5PFs7 zVEy><74;{Z6E$XjQ#JO>eYf9g{Gi z|1qhv8dTc=0E-*jkYH+gI+()!M z%7EdAr>j338xo6dsbWfwJ3gUY8yI=p2_B2K9%hx`$imZ!>LL)a$uJY03y3S#RQ&5~YQiTAOgfB6YiBgIB+KA=L>aNCisNXeFv;YAz zrjV9%@5nr_QAxZ?Pgz#w?E^EXynFB|94&EjkyoMPA`*QRduZ%~1zBC6gKlVR>=HfC z0LI|vv96;T=vG&7st;IEJyec{LDO&C!qrQ~`UZeYDi!3Qp>I+gF?xmWAL;leXsuj< z63RZe+MgXl;ocIAuEPzbVx~E>ljs%=Q4g_nX>;X&SV;y$#OByQbL-N0HsKYZnE=2il4^(U;-JuEAR|4n%m54wXSw5dIB-+VVS7MJQjyM#uO>#Z#2DPL z(Jc+-hgN;?GR8xY*NRh);_RFUjA zzT$8ilZlUR*5Hb97;zu6oELn58wK8UU56HuOz}WP_dQx{&l+CIDpU!x>mfh(Y$D?U z>5Ri@fQoC8-MeIjqfx4;+Vj#TLch_chEO`3Q6XVvrny;U+ z-M2~Pktx`bj-myP2b(v2L;CrPQ zZ+z7gPR)Iy4pk{LwFRA*M)<1FKBZ5T9MMe{E%KNg^zJ|fxU}{YeXD)$tO+Dt!f#qL&W6Ayi5jY!WCCwl0rxJXF`fs6@ z^X2arK(&%1@p}wG$LAIV@LBm^0cGmfnvnfhk!OD-q4}hQgavsYzO~z^yXP+Y!Yy5* zBm#hvf|qWVhCKT*|9_Em{v}V%c5bWqRJ4MsOvPgn(P;L!4D`1Q^tTN3@5(@8pZ=DC z{+5CMmVy43f&Rajf&L|GiGKmOPXKK?#$A+tGsvi(kP`vJl-kCw3xeTLc;xPezEjU% z3?66*%FZ&QO>BxrVYu?ucg!AZp$F%a}9TWuQmJzwtY3;<^-V@|H*zMO6hL z_vtnSiF^O0zZUn;C1|p!U;Waxb-8gZi#dk|qAgxQk`l7cuvL8ok-zgPFRda~{>U`= FzX23JaCHCx literal 0 HcmV?d00001 -- GitLab