ѧ
JOURN AL OF SOFTWARE
1999ꡡ1010ڡVol.10No.101999



վϵͳǨƻ 
   
ժҪǨǹվϵͳʵָƽ⡢ϵͳԹܵҪֶ.һֻڽ/ͷϢ¼ĽǨƼ.ϢݿPVM(parallel virtual machine)֮ʵ,жû͸ֲԺáСʵּ򵥵ص.˼ʵӦƵġглؾָǨϵͳChaRM(checkpointing-based rollback recovery and migration system).
ؼʡǨ,վ,PVM,״̬.
ͼšTP393
Process Migration for the Network of Workstations
PEI Dan WANG Dong-sheng SHEN Mei-ming
(Department of Computer Science and Technology Tsinghua University Beijing 100084)
Abstract Process migration is an important tool on NOWs (network of workstations) for load-balancing and high availability. In this paper, the authors present a log-based approach to provide process migration for parallel applications. Because this approach is implemented on top of PVM (parallel virtual machine), it is transparent to users and portable. This approach has been used in the ChaRM (checkpointing-based rollback recovery and process migration system) system. The experiments show that the overhead is low.
Key words Process migration, network of workstations (NOW), PVM (parallel virtual machine), process state.
ڹվ(NOW)ϵͳ,ĸطֲں̶ܴӰϵͳִЧ.ǨƼ̬ܶظıϵͳĸطֲ,֧ϵͳ̬ƽ⡢ЧԴϵͳܺϵͳԵĹؼ.[14]ָ,ýǨ,(1) ڼ,԰ѽ̴ӸؽصĽǨƵؽĽ,ʵֶ̬ƽ;(2) ԳNOWеĿйվ,Ҫռ乤վԴʱʱǨ;(3) ڳʱĹ,ʹĳ˳Խϵͳά,ϵͳĿ;(4) ʹI/O豸ڵĽϽ,ͨ.
PVM(parallel virtual machine)һӦù㷺Ĳб̻,䱾ûֽ֧Ǩ.PVMͨſ϶书ܽ,һڼƵĲглؾָͽǨϵͳChaRM(checkpointing-based rollback recovery and migration system).ChaRMϵͳṩ˲м㻷µݴܺͽǨƹ,븺ƽȹ,ʵֶԹվϵͳܻԴ.
̵״ָ̬,Ϊ˱֤Ǩܹ֮ȷִϢ.״̬Ӧð̵ݶΡûջ,ӦPC״̬PSWջָSPݡļϢԼжϡźŵϢ.ڽǨʱ,Ǩƽ̲״̬,ֱͨUNIXsocketͨϵͳΪĿϴһǼܽ;Ǽܽsocketͨж״̬,ԼĽ̿ռؽ״̬.״̬Ϻ,ǨƽԶֹ,ɹǼ̴ܽ.
ڲϵͳ,뱣֤ǨƺԭŷʸǨƽ̶ӦĹǼܽ,ұ֤ЩǨǰǨƹзǨƽ̵Ϣܹȷ˳,ϢĶʧִд.
1ʧϢĴ
ڽǨƹǨƽǼܽŲͬļϢʧͼ1ʾ.Ǩƴt0ʱ̿ʼ,t1ʱ̽(̵t0t1ܲһ).MPǨƽ,SPMPĿ϶ӦĹǼܽ,NMPӦóеķǨƽ.Ǩƹ,MP,NMP.Ǩƽ,MP,SPMPӦóв.

ͼ1ǨƿɵϢʧ
ͼ1(a),NMPt1ʱ֮MP͵Ϣm1ҲշMP,SPҲ޷յ,Ϣm1ʧ.NMPt1ʱ֮ȴMPϢ,SPϢm2ڷMP,NMP޷,m2Ҳʧ.Ϊ˱ȫǨɵĽŵĸı¶ʧϢ,һֽ֧ǨƵϵͳһMP,SP֮ƥ.ֻϵͳάMPSP֮ŵƥ,ҪMPͨʱ,ͨƥ鵽ӦSP,ȻSPʵʵͨ,ͽm1m2Ķʧ.,ƥƲܴͼ1(b)еm3ͼ1(c)еm4ĶʧϢ.NMPt0t1ʱ֮MPm3ʱ,ǨƵµű仯޷ӳNMPʱƥ,m3MPǷSP;ʱMPѾֹͣ,t1ʱ֮,޷յm3,Ϣm3ʧ.ͼ1(c)еm4m5t0ʱ֮ǰѾͳȥ,ǽսδյϢ.m3,NMPm4ʱ֪ŵı仯,m4MPʧ.Ϣm5ڲƥƶĶʧϢ,t1ʱ֮,NMPͨѯƥȴշSPϢ,޷յMPϢm5,Ϣm5ʧ.жͬʱǨʱ,Ǩƽ֮ĶʧϢͼ1ʾ,Ǳ˴t0t1ʱ֮䲻ӦϢ,Բm3һϢ.
ֽ֧ǨƵϵͳ,¼άƥķ.(1) ֱÿûά.ַضĲл,ÿû̵ռ䶼˶,Ҫͬʱƥ,άƥݵΨһ.(2) ɲб̻ϵͳÿϵdaemon(pvmd)ά.ַǰһַܽʡռ,ƥ,Ҫ޸Ĳб̻ϵͳԴ,ֲԲ.(3) ÿһ,רŸάƥ,ӹ̵ܽͨ.ַҲضĲл,ڽ̼ͨƵпܳΪϵͳƿ.
m3,m4,m5ĶʧϢķһ.(1) ͬʽ.ǨƵ,NMP(ͣ),m3ϢͲ;ڴ״̬ǰ,MPNMPֱm4m5,ϢĶʧ.ַʽʵּ,ǨƻᵼӦóͣ,,Чʵ.(2) 첽ʽ.ַʽNMPǨƹм.Ϊ˱ⶪʧm3,m4m5һϢ,MPVM[2]ϵͳ˸ӵϢת.ַʽִЧʸ,ʵָ,Ҫ޸Ĳб̻ϵͳԴ.
2ChaRMϵͳǨ
ΪChaRMϵͳһµĽǨƻ,NMPǨƹм,ֻǨƿʼͽʱм򵥵Э.Ϊ֧ڲͬл֮Ŀֲ,ChaRMϵͳһרŵĹC-Manager̸ϵͳ.Ϊ˴ͼ1еļϢʧ,ChaRMϵͳȡPVMװʽƥ䡢ϢϺϢӳٷ͵ȼ.
(1) PVMķװ. ִPVMǰ,һЩĹChaRMϵͳĹ,ṩûĽӿԭPVMĽӿȫһ.ַװ˶PVMԴֱ޸,ʹChaRMĹû͸.
(2) ʽƥ. ChaRMÿû̿ռάһƥ,ʹÿͨʱȲѯƥٽͨ.Ǩƽǰ,¸̵ƥ.,Ӧó̴ֻ֪ʱ,ȫ֪ýǷǨ.ֻЧرͼ1еm1m2һϢĶʧ.
(3) Ϣϻ. ChaRMPVMͨŻƱFIFO彫һһ̵Ϣͨ.Ǩƹ,AB̷һľϢͲBϢ;B̵ȴոþϢ,յӦϢƱ浽ԤָĽջ.FIFO,BյA̵ľϢʱ,ƶϴA̵B̵ĵͨͨеϢѱB.ֻշϢ¼ͼ1еm4,m5Ķʧ,ͬʱҲ֤ȷϢ˳.
(4) Ϣӳٷͻ. Ϊ˱ͼ1еϢm3,ChaRMPVMϢͺ˷װ,ʹõNMPҪt0t1ʱ֮MPӦϢʱ,ķ,ǽϢ¼һӳٷϢ.ȵǨٸݸ¹ƥӳٷϢеӦϢͳȥ.ʵһֻڷϢ¼.
㷨1. ChaRMϵͳǨ㷨.
1. C-Manager: IFյǨƷźMIGRATE_SIG THEN 
MPǨźMIG_CHECK_SIG;
NMPǨЭźMIG_SYNC_SIG;
2. MP:IFյMIG_CHECK_SIG THEN 
ͣ,һSPͨŵsocketܵ;
C-ManagerPRT_MSG,֪ͨsocketĵַ˿ںŵϢ;
н̷;ϢRDY_MSG;
͵Ϣ,ֱյн̵RDY_MSG.
IF յRDY_MSG Ϣ(϶;ӦϢ) THEN
Ϣ浽ԤָĽջ;
NMP:  IFյMIG_SYNC_SIG THEN 
MP;ϢRDY_MSG;
м;
IF Ǩ֮ǰ,ҪMPӦϢ THEN 
Ϣ¼ӳٷϢ,ķ;
C-Manager:IFյMPPRT_MSGϢ THEN
ĿMPӦĹǼܽSP;
MPsocketַ˿ϢͨPRT_MSG͸Ǽܽ;
3. MP:IFѾյн̷RDY_MSG THEN
ȴӦSPsocket;
״̬,״̬ͨsocketǼܽ;
״̬Ϻ;
SP: IFյC-ManagerPRT_MSG THEN 
ӦMPsocket;
socketж״̬,ԼĽ̿ռؽ״̬;
ű仯ͨRJN_MSG֪ͨC-Manager;
4.C-Manager:IFյSPRJN_MSG THEN 
NMPٴηMIG_SYNC_SIG;
н̷LST_MSGڸ̸ƥ;
NMP:IFյMIG_SYNC_SIGź THEN
͵Ϣ,ֱյзMPRDY_MSG;
IFյRDY_MSG Ϣ(϶;ӦϢ) THEN
Ϣ浽ԤָĽջ;
C-ManagerLST_MSG,¸Եƥ;
µƥӳٷϢӳٷ͵Ϣͳȥ;
ִ;
SP: IFյC-ManagerLST_MSG THEN 
ƥ;
ִ;
㷨ĵ2,MP̷;Ϣ,NMPMP;Ȼ,MPн͵ӦϢϵ;NMPм,ڴʱͽϢ,Ϊڵ2н̼˴˷͵Ϣ϶,һƳٵ4ʹNMPϢѵĿС.ڵ4,NMPϢϺ,ӳٷ͵Ϣͳȥִ.ӳٷͻƴĸӿС,ԭ:һ,NMPִϢ󲻱սǷյԼ,ӦϢNMPִеĹؼ·,ֻƶNMPļֱӰ;һ,ӳٷ͵ϢǨϺͳȥ,MPֱǨϺżͽӦϢ,ΪʱȴշNMPӦϢִʱ.
ͬʽĽǨ,NMPĿMPĿ൱;㷨1,NMPǨƷʱǨƽʱһЩЭ,һֱڼ,俪ԶСͬʽ.첽ʽĽǨ,ȻNMPûԵЭ,ַʽĸӵϢתҲӦóдһĿ;㷨1NMPǨƹвЭΪ,ʹ㷨ʵ,пС첽ʽ.,㷨1лڷ/շϢ¼㷨Чؽͬʽʵֺ첽ʽЧʸߵŵ.
3ܲ
ԻΪ4̨SUN Ultra2(ÿCPU,Ƶ200MHz)100M̫ӶɵĹվϵͳ,ϵͳΪSolaris2.5.Ϊܼ͵ľݳ,óһ4ϸ,ܹ8.Խ1ʾ,״̬ռָͨsocket͵Ľ״̬Ϣܺ;ʱָMPԴͣ㵽Ŀϼѵʱ,״̬ʱЭʱ;MPЭʱָϽн״̬,еѵʱ,аͼվϢC-ManagerͨšǼ̵ܽάƥȹ;NMPЭʱָNMPΪǨѵʱеֵ,ͼվϢC-ManagerͨšӳٷϢάƥȹ.
1 ݳǨƿ

״̬ռ
(MBytes)
ʱ(s)״̬
ʱ(s)Э
ʱ(s)NMPЭʱ(ms)
1.680.4180.2190.1892.0
2.000.4380.2530.1851.7
3.010.5840.3680.1962.6
4.000.6610.4720.1892.5
5.000.7590.5650.1943.5
6.010.9380.6840.2545.8
7.011.0390.8190.2207.5
8.021.1200.9160.2042.6



ӱ1Կ,״̬ʱǹʱ岿,Ž״̬ռ,³,²紫Ĳ,ⲿֿĴСҪȡӲ;ЭʱŽ״̬ռȻ,仯,0.2s,Э״̬ռĹϵС.ͼ2пԸֱ۵ؿMPֿ״̬ռĹϵ.1еNMPЭʱС,MPʱ,֤Ƕ㷨1ķ:NMPԶСͬʽ,ú뼶Эʵָӵ첽ʽǨҲȫԽܵ.һ𽥱ڴϴĽ״̬ռ䵼ͨű. 

ͼ2ݳMPǨƿͼ
,ǻһЩ˲,2.ƫ΢ַ̡任Ϊܼ,ASEάFFTΪܼͨ.Կ,ܼܼͨMPЭʱС,֤ϢϺӳٷͻƵЧʽϸ;ܼͨ͵Ӧó,NMPҪطϢȼܼ͵Ķ,Эʱһ,ͬʱһҲ״̬ռСӰ.
2 ӦóǨƿ

Ӧó״̬ռ
(MBytes)
ʱ(s)״̬
ʱ(s)Э
ʱ(s)NMPЭ
ʱ(ms)
任1.60.4380.2120.2152.0
ƫ΢ַ2.20.4710.2730.1963.1
ά1.70.4370.2250.2043.3
ASE8.01.0880.8640.2231.9
FFT3.70.4450.6790.2344.4
FFT22.12.2322.4450.2116.2
FFT66.17.3487.6870.2497.6

 
4ع
Ǩಢϵͳеõֺ֧Ӧ.ڵϵͳCharlotte,V system,Mosix,SpriteMach,ͨ޸Ĳϵͳں,ϵͳʵ˶ԽǨƵ֧[1].ַЧʸ,͸Ժ,ֲԲ.ûʵֵĽǨֻʹñ׼UNIXϵͳ,ֲԺ.MPVM[2]޸PVMԴ,˸ӵϢת֧첽Ǩ;Condor[3,4]ֵ֧ĽǨȽ̵ļļֹ̲ý,ȵϵͳҵлڸûϻָ;CoCheck[5]Condorṩĵ̼ú;ػָ,PVMͨſ֮ʵ,ֻ֧ͬʽĽǨ.DPVM[6]ýǨƽһ֧˶̬ȵĹ.
ϵͬ๤,ChaRMϵͳһЩܺص.(1) û͸,ʱķʽṩû.ûֻ轫ChaRMṩĿԭPVMӦó,ʽAPIýǨƹ.(2) ڲб̻ͨſʵ,רŵĹC-ManagerѡPVMԴRM,ӶضĲб̻,ֲԺ.(3) ͨϢϺϢӳٷͻ,֧һµġ/Ϣ¼ǨƼǨЧ,δøӵϢת.(4) ֹ֧ԶȫԶ3ַʽĽǨƷʽ.ûָĸǨĸ,ҲĳЩ̵Ǩ,ϵͳǨĸ,ϵͳоǨƷʱǨƵĽԼĿĽ.(5) Խ״̬˾,ųݶвر,ӶǨ״̬񡢴ؽʱտ.
ĽǨƼPVM,MPIȶϢݿʵ,ʽӳ䡢ϢϺϢӳٷͼ,һֻڽ/ͷϢ¼ǨƼ,ʵּ򵥡Ӧó͸ֲԺúͿСص.
һĹ,ǽChaRMϵͳƲʵֶ̬ƽ㷨,ԹվϵͳܻԴ.
עͣоõ863߿ƼĿ
߼飺ᵤ1973,˶ʿ,ҪоΪ/ֲϵͳ,ݴ븺ƽ
:1966,ʿ,,ҪоΪд,ݴ
:Ů,1938,,ʿʦ,ҪоΪ/ֲϵͳ
ߵλ:廪ѧѧ뼼ϵ  100084
E-mail: peidan@mail.cic.tsinghua.edu.cn
ο
1Eskicioglu M R. Design issues of process migration facilities in distributed 
systems. IEEE Technical Committee on Operating Systems Newsletter, 1989,4(2):3
13
2Casas J, Clark D L et al. MPVM: a migration transparent version of PVM. Computing
Systems, 1995,8(2):171216
3Litzkow J et al. Condor a hunter of idle workstations. In: Proceedings of the 8th
IEEE International Conference on Distributed Computing Systems. Los Alamitos, 
CA: IEEE Computer Society Press, 1988. 104111
4Tannenbaum T, Litzkow M. The condor distributed processing system. Dr. Dobbs 
Journal, 1995,(2):4048
5Stellner G, Pruyne J. Resource management and checkpointing for PVM. In: 
Proceedings of the 2nd European PVM Users Group Meeting. Lyon, France: Edition
Hermes , 1995. 131136
6Ͼű,κ,߳.DPVM:֧ǨƺŶӵPVM.ѧ,1997,20(10):872
877
(Ju Jiu-bin, Wei Xiao-hui, Xu Gao-chao et al. DPVM: an enhanced PVM supporting 
task migration and queuing. Chinese Journal of Computers, 1997,20(10):872877)
ո:1998-07-20޸:1998-10-09
