user manual
Contents
1. 15 Appendix C The workload transforming function 17 Appendix D Request sequence and their definitions 18 iii 1 Introduction 1 Introduction Search is a search engine model which is used to evaluate datacenter and cloud computing systems Search v1 0 brings some simplicity in terms of installation deployment and monitoring Within this version we are offering Search with everything inside and ready to go Search consists of a search engine a workload generator and a comprehensive workload characterization tool DCAngel 1 1 Targeted Audience This document is targeting two types of audiences People who just want to use Search as a benchmark tool for evaluating their datacenter and cloud computing systems This is for those who will directly use the provided Search benchmark directly to deploy it on their cluster People who would like to modify the sources to fit their particular needs You could use modified Search to do workloads characteristics analysis add some functionality or replace a component with another one 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 Search before How to install Search version 1 0 is introduced in Section 3 for people who are not going to make any change to the provided Search How to build an applia2. 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 py exps2 select comment avg hpc basic stall_breakdown inst mix Scache 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 la self pe exps2 select reqs comment netbytes from all where app search Part two regs comment netbytes head 100000 Ffixed 1008s2i 20regs SoGou throughputreall 4090 9054326 Ihead 100000 fixed 100052i 20regs SoGou Part three throughputreall 4090 9054326 head 100000 fixed 100882128reqs SoGou throughputreall 6665 93762575 head 100000 fixed 1000s2i 20regs SoGou throughputreall 193224 978873 head 100000 fixed 100 s8i2 cycle reqs 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 P
3. merged write sectors write time Appendix B 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 DCAngel database table structure For the meaning of all following table s abbreviations users can go over Appendix A field path app comment regs duration host Field path host insts cpi br miss ratio br stall ratio icache stall ratio tb stall ratio dcaceh stall ratio I2cache st
4. start time and end time of the workload The search directory collect the search log the terms send to search engine and warm up log The hmon directory collects performance data of Search Server nodes Query Server gt gt Ir I Extended SQI O Figure 2 High Level Diagram of DCAngel Users can get data through a browser using DCAngel For this version the only browser we supported is FireFox First you should start the service by executing the following commands Enter the directory python lib fsh cd python lib fsh Start the service psh py port For the port we use 8002 as a example psh py 8002 And then you can visit DCAngel s browser port through the address do not forget the slash after fsh http Client node ip address port Search exp cmds fsh 7 3 Getting started The Search above is the location of Search v1 0 package self py exps2 select reqs comment xplot path host 100 search latency from exps natural join all events where app search self py exps2 select comment xplot 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 exps2 select regs comment netbytes from all where app search self py exps2 select comment wavg active duration from all group by consent
5. the rate of Init step K 1 e g search anno ratestep 20 5 20 cf reg Appendix D Request sequence and their definitions Request sequence name warmup reqs reqs SoGou reqs Abc regs Xyz freguency freguency freguency regs by lens SoGou Sorting reqs SoGou according to request term s length reqs by lens Abc Sorting reqs Abc according to request term s length regs by lens Xyz Sorting reqs Xyz according to request term s length All the request sequence file above can be found in Search search engine data directory
6. Search A benchmark for datacenter and Cloud computing USER S MANUAL October 7th 2011 Revision Sheet Revision Sheet Revision Description 07 10 2011 Search v1 0 first packaging Memu USER S MANUAL TABLE OF CONTENTS CONTENTS Es Intr dUCHON A a 1 V MAG E 1 1 2 Structure of the CO cid ti 1 132 F rther CHL EE 1 UE CU taras 2 2 1 QCOuck introducliol ii A ed 2 22 Availableimplementations IA 2 3 Getting SAC 3 SA e hen a E E E E E E E ES 3 3 2 PEQUES E 3 3 2 Linux Kernel Versi t E 3 E MEE E 4 3 2 3 Python EET 4 E EMEP Dn EE 4 3 2 5e Malplotlib i eere EE OR EE EE P FOR PO E ere 4 3 20 JAVA 4 3 2 7 CPI oic ieiunii E MP SDN AN ar SEN 4 3 2 8 NNlcM ETT 4 3 2 9 Setup passphraseless Ssh unn e deese ta n p Fee ed EET NUI SE 5 3 210 lEE ev sto sde 5 So Deploying E EE 6 33 1 COMMPUPATION EET 6 3 4 Running Benchmark TN 6 341 Workload Preparation ido 6 3 4 2 Start benchmark t68t aede ene I Ee EES 6 9 A EE rb educi E m no Him ee one re ve 7 4 Building your own Search eee 0000000000000 onen none 10 dli vi e RE 10 4 2 Make yo rs arch engine cys ste vss diss die Sch antics eal vo Boek ees aoe Soo 11 4 3 Creating your own workload ss 11 Appendix A Metrics collected by DCAngel 13 Appendix B DCAngel database table structure
7. _stalls 194 RESOURCE STALLS BR MISS CLEAR 10dc br miss stalls 195 196 BUS TRANS ANY e070 bus trans Appendix B 197 BUS DRDY CLOCKS 2062 bus drdy 198 BUS BNR DRV 2061 bus bnr 199 BUS TRANS BRD e065 bus trans brd 200 BUS TRANS RFO e066 bus trans ro 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 Make your search engine For default Search we just supply a SoGou corpus s snapshot and indices and all the Search Server nodes have the same indices and snapshot it also called segments in nutch Your can use your corpus s snapshot and indices With your snapshot and indices you can separate the snapshot and index them by using the nutch command merge and index You should put each part of snapshot and index into Search Server nodes home ans42 crawl combinations directory The default Search gives you an example of the indices and snapshot s layout in each Search Server node s directory home ans42 crawl combinations After that you should modify the configuration file s i2 cfg in Cline node s Search nutch where represents the number of Search Server nodes The content of that configuration file is as follows 1 server list gd87 gd88 gd89 gd90 2 gd87 crawl dir 01 3 gd88 crawl dir 23 4 gd89 crawl dir 45 5 gd90 crawl dir 67 The first line represents the Search Servers hostnames From the second line each
8. all ratio res stall ratio rob stall ratio rs stall ratio Idst stall ratio fpcw stall ratio br mix load mix store mix Idst mix simd mix fp mix other mix bus util bus d util bus bnr ratio bus brd ratio bus rfo ratio Table exps Definition The test performance data s path under exp directory User used application s name The comment when user used to specify a Request name The test s duration Node s host name Table _all Definition The test performance data s path under exp directory Node s host name The mean value of instruction number Cycles per instruction Branch miss ratio Branch stall ratio Icache stall ratio TLB stall ratio Deache stall ratio L2 Cache stall ratio Resource related stall ratio Reorder buffer stall ratio Reserve station stall ratio Load and store stall ratio Float point unit stall ratio Branch instruction ratio Load instruction ratio Store instruction ratio Load and store instruction ratio SIMD instruction ratio Float point instruction ratio Instructions that except load and store ratio Bus utilization bus_drdy ratio Cusers can find bus_drdy and all the following abbreviations meaning in Appendix A bus_bnr ratio bus_brd ratio bus rfo ratio Appendix B cpu usage CPU utilization search latency Average guery latency search start Test start time duration The test s duration netbytes rnetbytes snetbytes netpackets rnetpacket snetpacket The meaning of following field is
9. ample 25872 means that we use s8i2 cfg as Search Server nodes configuration file where s8i2 cfg is in Client node s Search nutch directory reqfile indicates the original request sequence we use The request sequence file is in Client node s Search search engine data directory Appendix D lists the request sequence we have provided and users can use one of them or a new one In the example Oregs sogou means that we use sogou request and the request file is Search search engine data regs sogou You can use all the function in Appendix C to create your own workload and adopt your own Search Server nodes configuration file and request For how to configure Search Server nodes you can consult section 4 2 Appendix B 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 instruction
10. art 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 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 reqs comment netbytes from all where appz 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 arg1 standard deviation of argl corrcoef arg1 arg2 correlation coefficient between argl and arg2 correlate arg 1 arg2 cross correlation of argl and arg2 wavg argl arg2 weighted average of argl and arg2 is weight xplot arg1 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 arg and arg2 should be path and host respective arg3 is degree of data aggregation If arg3 equals 100 each 8 3 Getting started point in the figure represents the average value of 100 arg4 xhist argl arg2 a
11. ch Server 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 SHOME ssh amp amp In s authorized_keys2 authorized keys On the Client node test the results by ssh ing to other nodes ssh i HOME ssh id_dsa server This allows ssh access to the nodes without having to specify the path to the id dsa file as an argument to ssh each time 3 2 10 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 same network with regard to hardware and software configuration for example 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 address and the corresponding hostname 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 Getting started 3 3 Deploying Search You re suggested creat
12. defines the directory name of corresponding Search Server node s snapshot and index 4 3 Creating your own workload Section 3 4 1 mentions you can create your own workload and this section will explains how to create a workload Now we will show how to create a workload by show the syntax and explaining a given workload s meaning The given workload is as follows Syntax search anno function args function2 args configfile regfile An example search instance head 10000 poisson 20 s8i2 reqs sogou 11 Appendix B search means that a search engine is under evaluation We use dot to link different parts anno is the annotation of this workload in the example we use instance to indicate that this workload is an instance function args function2 args indicates the functions we use to the real request sequence function and function is transforming function s name The function can be found at Appendix C args is the function s parameters we use to link transforming functions In the example head 10000 means that we use head function in Appendix C head function s parameter is 10000 poisson 20 means that we use poisson function in Appendix C and its parameter is 20 configfile indicates the configuration file we used for Search Server The configuration file is in Client node s Search nutch directory In the ex
13. ests from clients and dispatching them to Search Servers We use Apache Tomcat 6 0 26 as the front end and nutch 1 1 as the search engine Search Server serving client requests transmitting by Web Server and the return the results to Web Server 3 2 Prerequisites The provided Search v1 0 relies on perf JDK Python and Numpy In this part we focus on how you can use what is provided in the Search v1 0 package for deeper information you may go over the Building part in section 4 Tomcat 6 0 26 and nutch 1 1 are included in our package so the user should not prepare them 3 2 1 Linux Kernel Version For this step you need to get the root privileges for your Linux servers We need to build a linux kernel whose version is 2 6 31 or newer for all the Search Server nodes because those kernels support perf events port which is used by perf When you compare the kernel you should make sure that perf events is build 3 3 Getting started into your 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 Search Server nodes and then enter the directory tools perf After that users 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 Client node needs Numpy http numpy scipy org w
14. 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 lowait 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 write
15. hich 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 Client node needs matplotlib http matplotlib sourceforge net which is a python 2D plotting library 3 2 6 JAVA Java 1 6 x preferably from Sun must be installed in all linux systems except Client node You should also set JAVA_HOME to the ans42 user 3 2 7 CPU For this version the Search Server 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 8 SSH SSH must be installed and sshd must be running To run the Search scripts that manage remote daemons please make sure that you can ssh on remote nodes without 4 3 Getting started entering password 3 2 9 Setup passphraseless ssh Client node must ssh to Web server and Search Server 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 HOME ssh id_dsa pub public key Copy HOME ssh id dsa pub to Web Server nodes and Sear
16. ing a new user for all Linux systems and use the new user to do the following 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 Search v1 0 package to the Client node using the user ans42 We assume that you put the decompressed package in the directory of Search All the following operations should be done in Client node 3 3 1 Configuration To deploy Search you should first configure the Search common mk file as follow uname ans42 the user s name for the benchmark upwd a the corresponding password of the user Master gd88 the Web Server node s hostname Node gd48 gd49 gd88 the hostname of Web Server node and Search Server nodes Do not change other configurations in this file At last execute make deploy and source bashrc Then Search will be deployed on all nodes The deployment time depends on the number of nodes and the machine s hardware configuration It maybe needs tens of minutes Before you running the benchmark please make sure that the Web Server node s port 9090 is available or the Web Server node s firewall has already been closed 3 4 Running Benchmark 3 4 1 Workload Preparation Enter the Search exp directory and edit the run test sh file 12 report search example head 100000 fixed 100 95712 reqs SoGou Here we give an example of workload at line 12 which is also a default workl
17. nce on your own needs can be found in Section 4 for people who are going to modify some components of Search 1 3 Further Readings The following links give more in depth details about technologies used in Search v1 0 Nutch http nutch apache org Perf https perf wiki kernel org index php Main Page Tomcat http tomcat apache org Sqlite3 http www sqlite org 1 Introduction Numpy http numpy scipy org Matplotlib http matplotlib sourceforge net 2 Search 2 Search 2 1 Ouick introduction Search is a search engine site benchmark that implements the core functionality of a search engine site providing indices and snapshot for a guery term It does not implement complementary services like crawling and ranking It only has one kind of session user s session via which users can query terms Search consists of three parts a search engine a workload generator and DCAngel The search engine is based on nutch which is an open source web search software project For Search v1 0 we use nutch 1 1 as the search engine s platform The indices and snapshot we used in Search are generated by nutch 1 1 with SoGou Chinese corpus http www sogou com labs dl t html We get a real world search engine s trace from a user s log of SoGou http www sogou com labs dl g html The workload generator can transform the real trace by specifying the query rate variation and terms situation The wo
18. nno head 100 0 cf req If start is O then is can be leaved out e g search anno head 100 cf reg unig NULL Get the unique query terms out of gs e g search anno uniq cf req random Total Randomly get query terms from gs and the total number of queried terms is Total e g search anno random 1000 cf req shuffle NULL Shuffle the terms in qs e g search anno shuffle cf req hot NULL Sort the gs according to the frequency of terms occurrence e g search anno hot cf req lens NULL Sort the qs according to terms length blockreq Blocksize repeatCount Repeat every Blocksize terms in qs RepeatCount times e g search anno blockreq 10 2 9 cf 9 req fixed Rate Generate fs and set the query rate to be Rate gueries per second e g search anno fixed 20 cf reg burst Rate K Generate ts and let ts be i K K Rate where i 1 len gs e g search anno burst 20 2 cf req scale Rate Compress or amplify original ts by setting the query rate to be Rate queries per second e g search anno scale 20 cf reg poisson Rate Generate ts and make the query rate variation fit poisson distribution and set the average rate to be Rate queries per second e g search anno poisson 40 cf req ratestep Init step K Generate ts and set the initial query rate to be Init The rate will increase for K 1 times Each time it will increase the value of 17 Appendix C step Finally it will be stable at
19. oad You can go over the workload part of session 4 if you want to create a new workload yourself If you want to use the default workload you should replace the 77 by the number of Search Server nodes 3 4 2 Start benchmark test Under the Search exp directory you should run the following command to start the 6 3 Getting started benchmark test make test The information of the test can be seen at file nohup out 3 4 3 Get result We have integrated DCAngel which is a comprehensive workload characterization tool in our Search benchmark Now we can use it to collect performance date aggregate data and visualize data 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 the Search exp log workload directory The workload here represents the workload you use For example if you use the default workload the log can be find at exp log search example head 100000 fixed 1000 s i2 reqs SoGou where Wa represents the Search server nodes number In that directory there will be a file named exp report if the test of the workload finished The file is an empty file and the only usage is to tell the user that workload replay has finished The exp log file records the
20. rg3 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 and 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 arg1 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 args 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 Appendi
21. rkload generator can also replay the real or synthetic traces DCAngel is a comprehensive workload characterization 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 For further reading about Search please look at the following site http prof ncic ac cn DCBenchmarks 2 2 Available implementations You may find available information and descriptions about older Search 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 ideas 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 Search 3 1 Overview Our experiment platform is based on Nutch s distributed search engine which is a typical two tier web application It offers the following architecture Search Server gt Client Web Server D lt IS Client injecting the workload thanks to the workload generator written in python and collecting metric results by DCAngel Figure 1 Architecture of Search Web Server receiving HTTP requ
22. the same as it in Appendix A So we will not explain them here iowait ctx active pgfault pgmajfault active conn passive conn read write read sectors write sectors For table all we also define some macro which you can use to simplify your inputting Fro 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 Idst 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 l2cache miss ratio bus bus util bus d util bus bnr ratio bus brd ratio bus rfo ratio proc basic cpu usage iowait ctx active pgfault pgmajfault net disk Sproc selected hpc all proc all 16 Appendix C Appendix C The workload transforming function In the following table we use qs and ts represent query sequence and time sequence respectively Function parameters Definition name head Total start Get gs and ts from the sequence number of start and the total entry number of gs and ts is Total e g search a
23. x A Users can look up Appendix A and write your own DCAngel command Appendix B 4 Building your own Search If you want to build your own Search 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 Search Server nodes do not own a CPU whose type is one of the types we mentioned in section 3 2 6 you should modify line 167 to line 201 of file Search 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 L1I MISSES 81 174 LID REPL f45 175 L2 LINES IN ANY f024 176 177 PAGE WALKS CYCLES 20c 178 CYCLES L1I 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 Ac 186 SIMD INST RETIRED ANY 1fc7 187 FP COMP OPS EXE 10 188 189 RESOURCE STALLS ANY Vide 190 RESOURCE STALLS ROB FULL Idec I91RESOURCE STALLS RS FULL 2dc 192 RESOURCE STALLS LD ST 4dc 193 RESOURCE STALLS FPCW 8dc F cpu cycles F bus cycles insets itlb_misses dtlb_misses icache_misses dcache_misses I2cache misses F 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
Download Pdf Manuals
Related Search