Home

DCAngel

1. write merged write sectors write time 4 Building your own DCAngel The number of processes that is running The number of processes that is blocked Total memory Memory that is not used Size memory in buffer cache Memory that cache used Memory that once was swapped out but still in the swapfile Memory that has been used more recently Memory that is not active Total amount of physical swap memory Total amount of free swap memory The number of pages that paged in from disk The number of pages that paged out to disk The number of page fault The number of major page faults TCP active connection TCP passive connection Received bytes Received packets Received error packets number Number of packets dropped by native network adapter Bytes sent Packets sent Number of error packets sent Number of packets dropped by remote network adapter Times of disk reads Times of disk merged reads Times of sectors read The total time disk read Times of disk writes Times of merged disk writes Times of sectors write The total time of disk write DCAngel collects those metrics per second and writes those values into exp log Appendix B Appendix B DC Angel database table structure For the abbreviations meaning in following table s users can refer to Appendix A Table exps field Definition path The test performance data s path under exp directory app User used application s name comment The comment when user
2. same network with regards to hardware and software configuration for example you may connect machines via a single hub or switch and configure the network interfaces to use a common network such as 192 168 0 x 24 To make it simple we will access machines using their hostname so you should write the IP addresses and corresponding hostnames into etc hosts The following is an example etc hosts 10 10 104 47 gd47 10 10 104 48 gd48 10 10 104 49 gd49 10 10 104 50 gd50 3 3 Deploying DCAngel You re suggested creating a new user for all Linux systems and log in with the new user to do the following step To make it simple we just assume the new user you created for the tool is ans42 with the password a The user should download the DCAngel v1 0 package to the Master node using the user ans42 We assume that you put the decompressed package in the directory of DCAngel All the following operations should be done in Master node 3 Getting started 3 3 1 Configuration To deploy DCAngel you should first configure the DCAngel common mk file as follow uname ans42 the user s name for the tool upwd a the corresponding password of the user node gd48 gd49 2d88 the hostname of Slave nodes Do not change other configurations in this file At last execute make deploy and source bashrc Then DCAngel will be deployed on all nodes The deployment time depends on the number of nodes and the machin
3. structure 14 iii 1 Introduction 1 Introduction DCAngel is a comprehensive workload characterization tool It can collect performance metrics and then write them into database for further analysis and visualization DCAngel v1 0 package brings some simplicity in terms of installation deployment and monitoring 1 1 Targeted Audience This document is targeting two types of audiences People who just want to use DCAngel to collect performance metrics This is for those who will directly use the provided DCAngel directly to deploy it on their cluster People who would like to modify the sources to fit their particular needs 1 2 Structure of the document This document goes on the following route A detailed introduction will be given in Section 2 for people who have never used DCAngel before How to install DCAngel v 1 0 is introduced in Section 3 for people who are not going to make any change to the provided DCAngel How to build an appliance on your own needs can be found in Section 4 1 3 Further Readings The following links give more in depth details about technologies used in DCAngel v1 0 Perf https perf wiki kernel org index php Main Page Sqlite3 http www sqlite org Numpy http numpy scipy org Matplotlib http matplotlib sourceforge net 2 DCAngel 2 DCAngel 2 1 Quick introduction DCAngel is a comprehensive workload characterizat
4. systems We need to build a linux kernel whose version is 2 6 31 or newer for all the Slave nodes because those kernels support perf events port which is used by perf Make sure that perf events options are selected in the configuration before compiling the kernel 3 2 2 perf For perf users should get a linux kernel source code whose version is 2 6 31 or newer on all Slave nodes and then enter the directory tools perf After that users 3 3 Getting started should execute the following commands to install perf make make install 3 2 3 Python All the linux systems need Python whose version is 2 7 Older or newer versions haven t been verified in our system 3 2 4 Numpy The Master node needs Numpy http numpy scipy org which is the fundamental package needed for scientific computing with Python You may need the following libraries or tools before installing Numpy atlas python nose lapack blas libgfortran python dateutil python matplotlib python tz python setuptools 3 2 5 Matplotlib The Master node needs matplotlib http matplotlib sourceforge net which is a python 2D plotting library 3 2 6 CPU For this version the Slave nodes CPU type must be as below 1 Intel Xeon processor 3000 3200 5100 5300 series 2 Intel Core 2 duo processor If you use other CPUs you may go over the CPU part in section 4 3 2 7 SSH SSH must be installed and sshd must be running To run
5. used to specify a regs Request name duration The test s duration host Node s host name Table _all Field Definition path The test performance data s path under exp directory host Node s host name insts The mean value of instruction number cpi Cycles per instruction br_miss_ratio Branch miss ratio br_stall_ratio Branch stall ratio icache_stall_ratio Icache stall ratio tlb_stall_ratio TLB stall ratio dcaceh_stall_ratio Deache stall ratio I2cache_stall_ratio L2 Cache stall ratio res_stall_ratio Resource related stall ratio rob_stall_ratio Reorder buffer stall ratio rs_stall_ratio Reserve station stall ratio Idst_stall_ratio Load and store stall ratio fpcw_stall_ratio Float point unit stall ratio br_mix Branch instruction ratio load_mix Load instruction ratio store_mix Store instruction ratio Idst_mix Load and store instruction ratio simd_mix SIMD instruction ratio fp_mix Float point instruction ratio other_mix Instructions that except load and store ratio bus_util Bus utilization bus_d_util bus_drdy ratio users can find bus_drdy and all the following abbreviations meaning in Appendix A bus_bnr_ratio bus_bnr ratio bus_brd_ratio bus_brd ratio bus_rfo_ratio bus_rfo_ratio Appendix B cpu_usage CPU utilization duration The test s duration netbytes rnetbytes snetbytes netpackets rnetpacket snetpacket The meaning of following field is the same as it in Appendix A So we will not explain them here iowait ctx active pg
6. DCAngel A comprehensive workload characterization Tool USER S MANUAL October 7th 2011 Revision Sheet Revision Sheet Revision Description 07 10 2011 DCAngel v1 0 first packaging Memu USER S MANUAL TABLE OF CONTENTS CONTENTS I INH uni CE HE EIE 1 dd Tarseted ANNES NR E Emm 1 1 2 Structure of the dE nn 1 T5 VEN Readings od asc tesa Rd KO N ORE VERSEHEN 1 De DE AND EE EE ee 2 2 1 MEE AMON aaa 2 22 Availableimplementations zei 2 3 Getting SATION NR 3 Sl MN Wis een AE EE EE DE EE 3 32 AE EE IE N EE EE N EE EN 3 3 21 Linux Kernel SEE ee e abe en an 3 UMEN 3 3 2 3 PYON EA 4 324 Numpy i eee e eto tee rri baee rine t nen expe ee enden dan 4 3 2 5 EYE EE EE een coca N EE ended 4 3 2 6 CPU EE EE EO EE RE EE EE 4 3 2 7 Nic E 4 3 2 8 Setup passphraseless Ss icai eode er RR ERG ge GED e EE ees 4 3 2 9 es AE uam tee E ne RAS UE eus OR 5 3 3 BIO Da DE MBL Dr cmm 5 33 1 slui lo EA ER N 6 3 4 Running DE Anel RR AE Ee DE OG De etende 6 3 4 1 Add your own application cei see ee ee ee ee ee Re ee ee ee ee Re ee ee ee ee 6 342 Start DCAngel eee se E UE GEE Ge Ee E Ge Sees 6 OMEN HIT EE EE EE EE AR ER a EE 7 4 Building your own DCAngel ee ss se ee es ee eene enne 10 MEO UC EE OE OE OE ae 10 2 25 9eDarale DE NAVE SES GE ODE DES ayy de Rena SG DE SG 11 Appendix A Metrics collected by DCAngel 12 Appendix B DCAngel database table
7. e s hardware configuration It maybe needs a few minutes 3 4 Running DCAngel 3 4 1 Add your own application Enter the DCAngel exp directory and edit the Makefile file 54 THHHHHHHHHHHHBHIstart your application here THHEHHHHHHHHHHHHHHHHHHHHHHHHBE 55 application 56 echo application start 57 sleep 5 58 59 THHHHBHHBHHHHHHBHHBHHBHHHHHBHHBHHHHHHHBHHBHHBHHHHHBHHBHHHHHHBHHBHBHBHHHBHHBHBHE You should add your application command from line 56 The application here is sleep and the DCAngel will collect Slave nodes metric data when Master node sleep for 5 seconds 3 4 2 Start DCAngel Enter the DCAngel exp directory and edit the run test sh file 12 report app annotation workload The string of app annotation workload is the directory s name for your application s performance data You should replace app annotation workload with your own words But you must make your command in the same format as above Note that we use a dot to connect two sections app means your applications name annotation is the annotation of this test workload means the information or characterization of this application s workload Those strings are just used to 6 3 Getting started distinguish different test and users can use any words they like For the sleep application you can write following statement to run test sh file report sleep 5seconds NULL sleep means that your appl
8. fault pgmajfault active_conn passive_conn read write read sectors write sectors For table all we also define some macros which you can use to simplify your inputting For example you can write a DCAngel command self py exps2 select prim from all which has the same function with self py exps2 select app comment reqs host from _all Macros and their definitions macros prim app comment regs host hpc_basic insts cpi br_miss_ratio stall_breakdown br_stall_ratio icache_stall_ratio tlb stall ratio dcache stall ratio I2cache stall ratio res stall ratio rob stall ratio rs stall ratio ldst stall ratio fpcw stall ratio inst mix br mix load mix store mix ldst mix simd mix fp mix other mix cache itlb miss ratio dtlb miss ratio icache miss ratio dcache miss ratio I2cache miss ratio bus bus_util bus_d_util bus_bnr_ratio bus_brd_ratio bus_rfo_ratio proc basic net active conn passive conn netbytes netpackets Sdisk proc_selected Shpc all Sproc all 15
9. ication s name is sleep the annotation Sseconds means it sleeping for 5 seconds and NULL means this workload dose nothing in this 5seconds After that you should run the following command under DCAngel exp directory to start the test and let DCAngel collect performance data make test 3 4 3 Get result After your application running over you can use DCAngel to get performance date aggregate data and visualize date Figure 2 shows the high level diagram of DCAngel It stores performance data in a relational database managed by SQLite3 that supports the extended SQL statements Users can access those data through the extended SQL statements All the tests log and performance data collected by DCAngel can be find in DCAngel exp log app annotation workload directory where the app annotation workload is part of the make command you input to start a test For example if you use the make command in section 3 4 2 the log can be found at DCAngel exp log sleep 5second NULL In that directory there will be a file named exp report if the test have finished The file is always empty as it is just a signal to tell users whether the whole test has finished The exp log file records the start time and end time of the test The Amon directory collects performance data of Slave nodes We have already included an example test result in the DCAngel package which you can find at DCAngel exp log sleep GivenExam
10. ion tool It can collect performance metrics and then write them into database for further analysis and visualization We use perf to collect performance counters data and sqlite3 as the database The whole tool is developed in python 2 2 Available implementations You may find available information and descriptions about older DCAngel versions at its home page http prof ncic ac cn DCBenchmarks If newer version implemented it will be appended If you find some bugs please contact us via jiazhen ncic ac cn If you successfully implement it on your own platform please let us know If you have some novel idea you might share with us 3 Getting started 3 Getting started In this part you will drive right into the configuration and running part supposing you don t want to modify the provided DCAngel 3 1 Overview DCAngel adopts a simple master slaves model as shown in Figure 1 Slave ze d a Figure 1 Architecture of DCAngel Master starting your application and controlling Slaves to collect metric data Slave collecting metric data and sending those data to Master 3 2 Prerequisites The provided DCAngel v1 0 relies on perf and Python In this part we focus on how you can use what is provided in the DCAnge v1 0 package for deeper information you may refer to the Building part in section 4 3 2 1 Linux Kernel Version For this step you need to get the root privileges for your Linux
11. ld your own DCAngel this part will give some advices If following introductions do not suffice for your approach you may contact us via jiazhen ncic ac cn 4 1 CPU If your Slave nodes do not own a CPU whose type is one of the types we mentioned at section 3 2 6 you should modify line 167 to line 201 of DCAngel hmon hmon py 167 kperf events map 168 CPU CLK UNHALTED CORE 3c 169 CPU CLK UNHALTED BUS 13c 170 INST RETIRED ANY cO 171 ITLB MISS RETIRED c9 172 DTLB MISSES ANY 108 173 LIT MISSES 81 174 LID REPL f45 175 L2 LINES IN ANY f024 176 177 PAGE WALKS CYCLES 20c 178 CYCLES LIT MEM STALLED 86 179 180 BR_INST_RETIRED ANY c4 181 BR_INST_RETIRED MISPRED c5 182 183 INST_RETIRED LOADS 1c0 184 INST RETIRED STORES 2c0 185 INST RETIRED OTHER 4c0 186 SIMD INST RETIRED ANY 1fc7 187 FP COMP OPS EXE 10 188 189 RESOURCE STALLS ANY Ifdc 190 RESOURCE STALLS ROB FULL Idc I91RESOURCE STALLS RS FULL 2dc 192 RESOURCE STALLS LD ST 4dc 193 RESOURCE STALLS FPCW 8dc cpu cycles bus cycles insets itlb_misses dtlb_misses icache_misses dcache_misses I2cache misses page walks icache stalls br insts br misses load insts store insts other insts simd insts fp_insts res_stalls rob_stalls rs stalls Idst_stalls fpcw_stalls 194 RESOURCE_STALLS BR_MISS_CLEAR 10dc br miss stalls 195 196 BUS TRANS ANY e070 197 BUS DRDY CLOCKS 2062 bus trans bu
12. nd arg2 should be path and host respective arg3 is degree of data aggregation If arg3 equals 100 each value on the x axis represents the average value of 100 arg4 xscatter argl arg2 arg3 arg4 arg5 draw bi dimensional histogram of arg4 and arg5 argl and arg2 should be path and host respective arg3 is degree of data aggregation If arg3 equals 100 each value on x axis and y axis represents the average value of 100 arg4 and arg5 xcorr argl arg2 arg3 arg4 arg5 plot the cross correlation between arg4 and arg5 argl and arg2 should be path and host respective arg3 is degree of data aggregation If you want to use xplot you must make sure that the following read color words are not changed self py exps2 select reqs comment host xplot path host 1 metric from exps natural join all_events self py exps2 select reqs comment host xhist path host 1 metric from exps natural join all events self py exps2 select reqs comment host xscatter path host 1 metric metic from exps natural join all events self py exps2 select reqs comment host xcorr path host 1 metric metric from exps natural join all events For metric it can be any metircs can be any field in Appendix B We list the table structure of DCAngel s database in Appendix A Users can look up Appendix A and write your own DCAngel command 4 Building your own DCAngel 4 Building your own DCAngel If you want to bui
13. nstruction fetches stalled Retired branch instructions Retired mispredicted branch instructions Instructions retired which contain a load Instructions retired which contain a store Instructions retired which no load or store operation Retired Streaming SIMD instructions Floating point computational micro ops executed Resource related stalls Cycles during which the reorder buffer full Cycles during which the reserve station full Cycles during which the pipeline has exceeded load or store limit or waiting to commit all stores Cycles stalled due to floating point unit control word writes Cycles stalled due to branch misprediction All bus transactions Bus cycles when data is sent on the bus Number of Bus Not Ready signals asserted bus trans brd Burst read bus transactions bus trans rfo Read For Ownership bus transactions Metrics from proc filesystem usr nice Sys idle iowait irq softirq intr ctx procs User mode CPU time The CPU time of processes whose nice value is negative Kernel mode CPU time Idle time Iowait time Hard interrupt time Soft interrupt time The times of interrupt happened Context switch times Process number running blocked mem_total free buffers cached swap_cached active inactive swap_total swap_free pgin pgout pgfault pgmajfault active conn passive conn rbytes rpackets rerrs rdrop sbytes spackets serrs sdrop read read merged read sectors read time write
14. own commands A DCAngel command has two parts a fixed part and a SQL like part Let us look at the following command as an example self py exps2 select regs comment netbytes from all where app search The fixed part is self py exps2 and the SQL like part is select reqs comment netbytes from all where app search For the SQL like part users can write any statement that meets the sglite3 s syntax DCAngel s feedback may take a few seconds if it is your first time to execute a DCAngel command after a test That is because DCAngel needs time to write metrics data it collected into database DCAngel also defines many extend SQL functions Those functions usage are shown as below std argl standard deviation of argl corrcoef argl arg2 correlation coefficient between argl and arg2 3 Getting started correlate argl arg2 cross correlation of argl and arg2 wavg argl arg2 weighted average of argl and arg2 is weight xplot argl arg2 arg3 arg4 draw the scatter figure of arg4 The x axis of this figure is time and the y axis is arg4 s average value argl and arg2 should be path and host respective arg3 is degree of data aggregation If arg3 equals 100 each point in the figure represents the average value of 100 arg4 xhist argl arg2 arg3 arg4 draw the histogram of arg4 s occurrence times The x axis of this figure is occurrence times and the y axis is arg4 s average value arg a
15. ple NULL Query Server ee mm Dee E a I Extended SQI Figure 2 High Level Diagram of DCAngel We can get data through a browser using DCAngel For this version the only browser we supported is FireFox First we should start the service by executing the following commands enter the directory DCAngel python lib fsh cd DCAngel python lib fsh 7 3 Getting started Start the service psh py port For the port we use 8002 as an example psh py 8002 And then we can visit DCAngel s browser port through the address do not forget the slash after fsh http Client node ip address port 5DCAngel exp cmds fsh The DCAngel above is the location of DCAngel v1 0 package self py exps2 select reqs comment zplot path host 100 search latency from exps natural join all events where app search self py exps2 select comment zplot path host 1 search latency from exps natural join all events self py exps2 select reqs comment search latency cpu usage read cpi insts inst mix stall breakdown from all where app search self py ezps2 select reqs comment netbytes from all where app search self py exps2 select comment vavg active duration from all group by comment Part one self py exps2 select comment wavg br icache tlb dcache l2cache res rob rs ldst stall ratio duration from all group by comment self
16. py exps2 select comment avg hpc_basic stall_breakdown inst_mix cache bus from exps natural join cpi corrcoef group by comment term txt self py exps2 select reqs comment xplot path host 1 proc all from exps natural join all events where app search self py exps2 select from cpi corrcoef natural join exps Ref This Cmd Output EN self py exps2 select reqs comment netbytes from all where app search Part two reqs comment netbytes head 100000 fixed 100 s2i 2 reas SoGou throughputreall 4090 9054326 head 100000 fixed 1008s2i 2 reqs SoGou Part three throughputreall 4090 9054326 head 100000 fixed 1008s2i2 regs SoGou throughputreall 6665 93762575 head 100000 fixed 1008s2i 2 reqs SoGou throughputreall 193224 978873 Ihead 100000 fixed 100 s8i2 cycle regs SoGou throughputreall 67895 3581801 Figure 3 snapshot of DCAngel s GUI Figure 3 shows the snapshot of DCAngel s GUI The GUI can be divided into three parts Part one is commands column Each line in that column is a DCAngel command Users can execute the command by ctrl left mouse button click Users can edit those commands to meet your requirement Part two is command input column you can input your command here and execute it by pressing Enter Part three is a display column which displays the result of the command Now we will show you the DCAngel command s grammar so that you can writer your
17. s drdy 4 Building your own DCAngel 198 BUS_BNR_DRV 2061 bus_bnr 199 BUS_TRANS_BRD e065 bus_trans_brd 200 BUS_TRANS_RFO e066 bus_trans_rfo 201 t You should go over your CPU s software design manual and change hexadecimal number above to the corresponding CPU event number 4 2 Separate DCAngel You can use DCAngel only for collecting data or analysis data The metrics data DCAngel collected could be found in DCAngel exp log directory Users can use another tool to analyze data Users can also use another tool to collect metrics data and make sure that those data have the same format as DCAngel collected data Then users can visit DCAngel s browser port to get data and visualize data 11 4 Building your own DCAngel Appendix A Metrics collected by DCAngel variable Definition Metrics from performance counters cpu_cycles bus_cycles insts itlb misses dtlb misses icache misses dcache misses page walks icache stalls br insts br misses load insts store insts other insts simd insts fp insts res stalls rob stalls rs stalls Idst stalls fpcw stalls br miss stalls bus trans bus drdy bus bnr Core cycles when core is not halted Bus cycles when core is not halted Retired instructions Retired instructions that missed the ITLB Memory accesses that missed the DTLB Instruction Fetch Unit misses L1 data cache misses Duration of page walks in core cycles Cycles during which i
18. the DCAngel scripts that manage remote daemons please make sure that you can ssh on remote nodes without a passphrase 3 2 8 Setup passphraseless ssh Master node must ssh to Slave nodes without a passphrase Now check that ssh localhost If you cannot ssh to nodes without a passphrase execute the following commands at Client node ssh keygen t dsa f HOME ssh id_dsa P This should result in two files HOME ssh id_dsa private key and 4 3 Getting started HOME ssh id_dsa pub public key Copy HOME ssh id_dsa pub to Slave nodes On those nodes run the following commands cat id_dsa pub gt gt HOME ssh authorized_keys2 chmod 0600 HOME ssh authorized_keys2 Depending on the version of OpenSSH the following commands may also be required cat id_dsa pub gt gt HOME ssh authorized_keys chmod 0600 HOME ssh authorized_keys An alternative is to create a link from authorized keys2 to authorized keys cd HOME ssh amp amp In s authorized_keys2 authorized keys On the Master node test the results by ssh ing to other nodes ssh i HOME ssh id_dsa server This allows ssh to access to the nodes without specifying the path of the id dsa file as an argument to ssh each time 3 2 9 Network This should come as no surprise but for the sake of completeness we have to point out that all the machines must be able to reach each other over the network The easiest is to put all machines in the

DCAngel

Contents

Download Pdf Manuals

Related Search

Related Contents