Study Notes

Next week,I will give Oracle training to some people.
It has important meaning.

Modify the kernel to make Linux run faster:

# cd /usr/src/linux-2.4 
# make mrproper 
# make menuconfig

This step will display a dialog,and I exclude some items which will never be used, for example, Fiber channel,isdr,and pcmcia,(Type “N”),and then save the new configuration,issue the following command:

# make dep (it will take about 5 minutes) 
# make clean (5 seconds) 
# make bzImage (10 minutes) 
# make modules (it's long and will take about 30 minutes) 
# make modules_install

Note: it is “modules_install” which have no backspace between “modules” and “_install”. and if issue the “make modules -install”, the process will failed with error. The final step is issueing the “reboot”. I’ve tried all the above steps, and my Linux runs faster after rebooting.

nroff: emulate nroff command with groff.

# nroff -man man/libnet.3 | less

Issue the commands as other user account:

# su - username -c "ls"

For example, let root issues “ls” :

# su - root -c "ls"

Uncompress the “.bz” file:

# tar -jvxf xxx.bz

If the file “/etc/inittab” was modified, issue the following command to make the configuration activity.

# init q

Q or q, tell init to re-examine the /etc/inittab fill.
S or s, tell init to switch to single user mode.

Issue some continuous commands, and it will break as soon as error happens.

# command1 && command2 && command3 

Uncompress the file included in rpm:

# cat xxx.rpm | rpm2cpio | pax -r

How to mount harddisk and directory owned by local Windows system or Remote windows system?

# mount -t smbfs -o username=x, password=x, //share_path /mnt/c 

Ethereal — capture packets.

Create and use .iso file:

# mkisofs -J -r -o xxx.iso /test 
# mount -o loop xxx.iso /mnt/cdrom 

Uncompress xxx.bz file:

# tar -I xxx.bz or # bunzip2 xxx.bz

Check the file’s rpm name:

# rpm -qf /usr/bin/ls

List description information:

# rpm - qpi

List file information:

# rpm - qpl

Configure the Memory limitation in Grub:

grub> kernel /boot/vmlinuz-2.4.19 20 boot=/dev/hda2 mem=120MB 

Type Chinese output:

# ls --show-control-chars

Make a Linux boot disk:

# mkbootdisk --device /dev/fd0 2.4.20 

Where 2.4.20 is directory in /lib/modules

Modify the ethernet device’s MAC address:

# ifconfig eth0 down 
# ifconfig eth0 hw ether 00:AA:BB:CC:DD 
# ifconfig eth0 up 

A very important command which is used to check ethernet card’s statue:

# mii-tool -v 

mii-tool -view, manipulate media-independent interface status.This utility checks or sets the status of a network interface Media Independent Interface(MII) .
-w watch monitor for link status changes

# du ./ --max-depth=1 
# du -sm /directory

List the existed users account in the Linux:

# cat -fl -d: /etc/passwd 

Find the MAC address of local ethernet card:

# arp 
# /sbin/arp 

I have two computers in the office, the better one is installed RHEL3.0 update4. It used one Pentium4 775 3.2Ghz CPU and 4GB DDR2 memory,all these were based on Intel SE7221BA1-E engineer sample board. How fastly it runs! Another contains Celeron 533 and poorly 128MB memory,there are two operating system installed in this PC, windows 2000 Professional and Red Hat Linux 9.0.

The Lutos Notes is released by MI recently,and every employees must use this office system. I think the software is famous and be benifit for increasing work rate, but in facl it runs very very slowly so that I have no patience to wait. Through almost my work is done on Linux, I tried to control the Windows computer in my own Linux environment.

I found two tools, VNC and rdesktop. Both of these software were installed successfully, but it seemed that VNC could not response. Rdesktop would work but still could not connect to the windows. It said the windows refused connections. Why?

After reading some document in the Internet, I found that Windows 2000 Professional does not support remote connection ablity or terminal utility. However, Windows 2000 Server families could do this after installing external packages. Remote desktop was embeded in Windows XP and 2003 server. So, I installed a new Windows XP Professional in the PC whose cpu was celeron 533Mhz.

And then issue the rdesktop in the Linux command line:

# rdesktop -u phillip -g "1024*768" 192.168.1.143

“u” means username, phillip is an account of Windows XP
“g” configure the screen displayed in Linux environment
192.168.1.143, IP address of Windows machine.

By the way, the port number, for example 45272, could also be added to the following IP address.It will be useful when connecting to a server has firewall protecting. Everything of the Windows runs well in the Linux rdesktop connection session.

Often we want to execute parallel jobs consisting of standard UNIX commands.By “parallel” we mean the same command runs on multiple nodes of the cluster.Rocks provides a simple tools for this purpose called cluster-fork.

By default,cluster-fork uses a simple series of ssh connections to launch the task serially on every compute node in the cluster.
Cluster-fork is smart enough to ignore dead nodes.Usually the job is “blocking”,cluster-fork waits for the job to start on one node before moving to the next.
By using the “–bg” flag you can instruct cluster-fork to start the jobs in the background.

Often you wish to name the nodes your job is started on,this can be done by using SQL statement or by specifying the nodes using a specific shorthand.

1.frontend,use the SQL database on the frontend:

$ cluster-fork --query 
"select name from nodes where name like 'compute-1-%'" [cmd]

2.Second method
requires us to explicitly name each node:

--node=compute-0-%d:0-2  : compute-0-0,compute-0-1,compute-0-2
--node=compute-0-%d;0,2-3  : compute-0-0,compute-0-2,compute-0-3 

Compute nodes Installation: 1.Login to the frontend as root; 2.Run a program which capture a compute node DHCP requestes and put their information into the Rocks MySQL database:

# insert-ethers

If the frontend and compute nodes are connected via a managed ethernet switch,select “Ethernet Switches”.”Insert-ethers” is waiting for new compute nodes.Power up the first compute node,when the frontend machine receives the DHCP request from the compute node,it indicates that insert-ethers received the DHCP request from the compute node,inserted it into the database and update all configuration files.

When reinstalling very large clusters,load is managed on the frontend by telling compute nodes to back off and retry if the server is too busy.Specially,if the number of concurrently running kickstart.cgi processes exceeds some threshold,the fronted will send a “negative acknowledge” in the form of 503 http error signal.When a compute node receive this signal,it takes a reasonable action and retries its installation later.While the load threshold is reasonable for most clusters,it may be adjusted in the /opt/rocks/etc/kcgive file.Once adjusted,the value is used immediately.

Crossing kickstart:Contain nodes of different hardware architectures.

Mount the CD’s as ISO format:

# mount -o loop *.iso /mnt/cdrom

Frontend installation

Required the following CDs: Rocks Base CD,HPC Roll CD,Kernel Roll CD.As soon as boot off from CD,type “frontend”.This step must be operated fastly before the screen displays the “boot” prompt,or the CD will be told to install a compute node,not frontend we need. We could access the kickstart system by pressing Alt+F3 and Alt+F4 respectively.According to IA-64,the following command should be used:

$ frontend driverload=ide-scsi(or hdb=ide-scsi)

Installation steps:

1.Input CDs,and type “frontend”;
2.Input the cluster information;
3.Partition (Automatic or Disk Druid),If select manual partitioning,you must specify at least 6GB for the root partition and you must create a sperate /export partition.
The default configuration is automatic.
4.Private cluster network configuration;
5.Public cluster network configuration,Note,DHCP is not supported,we must specify the IP address.
6.Configure the gateway and DNS entries;
7.Set the timezone and NTP server;
8.Input the root password;

Then the frontend will format it’s file systems,and install the packages.

Customers can concentrate on their scientific computing instead of worrying about the management of their cluster.Stability,maintainability,ease of installation are the target of NPACI Rocks.

Complexity of cluster management,for example,determine if all nodes have a consistent set of software.In additional,this structure insures any upgrade will not interface with actively running jobs.

A cluster contains compute nodes,frontend nodes,and monitoring nodes.Within a distribution,different node types are defined with a machine specific Red Hat kickstart file,made from a Rocks Kickstart Graph.

A kickstart file is a text-based description of all the software packages and software configuration to be deployed on a node.I found Kickstart item in the Red Hat AS3.0 update4 which is the operating system of my office computer.It shows the information of the setup and configuration about hardware.

Rocks is robust system.Rocks makes complete operating system installation on a node the base management tool.With attention to complete automation of this process,it became FASTER to reinstall all nodes to a KNOWN CONFIGURATION than it is to determine if nodes were out of sychronization in the first place.So,it is a useful way to reinstall the system than determination the difference of nodes.

Rocks suppors all the hardware component that Red Hat support,but only support the following types hardware:x86,IA-64 architecture(No SPARC,Alpha or Yamhill):

Support:x86(IA32,AMD),IA-64(Itanium,Mckinley),x86-64(AMD opteron)
Network:Ethernet,Myrinet(Lanai 9x)

Physicial setup for Rocks:

1.Frontend

Nodes of this type are exposed to the outside world,many services run on there nodes.Frontend nodes are where users login in,submit jobs,and compute code,etc.This node can also act as router for other cluster nodes by using network address translation.Frontend nodes generally have the followiing characts:Two ethernte interface,one public,one private; Lots of disk to store file.

2.Compute

There area the workhorse nodes.They are also disposed. The complete OS to be reinstalled on every compute node in a short amount of time(~10min).There nodes are not seen on the public Internet.The main component:Power cable,Ethernet Connection for administration,Disk drive for caching the operating system(OS and libraries),optional high-performance network(Myrinet).

3.Ethernet Network

All compute nodes are connected with ethernet on the private network.This network is used for administration,monitoring,and basic file sharing.

4.Application Message Pass Network

All nodes can be connected with Gigabit-class network and required switches.

On the compute nodes,the Ethernet interface that Linux maps to eth0 must be connected to the cluster’s Ethernet switchs.This network is consided private,that is,all traffic on this network is physically separated from external public network.
On the frontend,two ethernet interface are required.The interface that Linux maps to eth0 must be connected to the same ethernet network as compute nodes.The interface that Linux maps to eth1 must be connected to external network.

Once you’ve physically assembled your cluster,each node needs to be set to boot without a keyboard.This procedure requires setting BIOS values and,unfortunataly,is different for every motherboard.We’ve seen some machines where you cannot set them to boot without a keyboard.

Paxville processor, made by Intel, would work as 4 processors because of its dual-core technology and hyperthread.This means one processor acts as two independent core and each core would be two processor by hyperthread technology. So we could say one Paxville processors equals four independent processors.

It took me very long time to choose two Paxville DP from six samples which have all same appearence. Before I used these identified Paxville DP, I even used Paxville MP in the SE7520JR2 mainboard,it seems crazy that four status laps glare in red color and the whole machine was halted. Then reflashed the mainboard’s BIOS to P09 reversion,and used Paxville DP,everything was ok.

There’s another issue,when I put the mainboard with two Paxville DP and four Samsung memory into the SR1450 chassis,I found if I used the line connects the front console panel with mainboard by passing the back panel,the system was halted. Howerver,I tried to connect the front panel with mainboard directly,without passing through back panel, the machine started up well with no error reporting. I’ve wrotten to James and Lily for Intel’s support.

The engineer and sales of Shanghai said our platform including PR1510 and PR1325 had a very bad performance.PR1325 used intel’s chassis named SR1425 with SE7320VP2.The customer wanted to test this machines by SuSE server 9.0. From the report created by Shanghai’s sales, I think perhaps it is caused by Hitach’s harddisk. I’ve suggested the customer to configure the /Swap partition to twice of total memory’s capacity. This issue is still open and I’ll resolve it in future.

In the afternoon,a customer asked me to modify the boot sequence controlled by lilo. I modified /etc/lilo.conf file,and fix this problem successfully.

Chris and Jin leaved Powerleader in 15th,September.They are both very good men from which I’ve learn many things,especially carefulness and responsiblilty. I think I would work with them for more years,but it isn’t in the same way.

Summary of last two days:
——————————————-
1.Mr Ji, from Xiamen City,said PR2510 could use many cpu resource on the Red Hat Advanced Server 3.0 with Update 4 even there was no process running. He also told me the avg cpu time is very high when he tried to copy some big file to local harddisk. I’ve read his report which seems have heavy load created by issued iostat command.
In the Saturday morning I installed the RHAS3.0 u4 in a SE7520BD2 platform,and then issued the top and ” # iostat -k 1 9 ” command, the result I got was normal.it showed the cpu works very well in our products.
I’ve written to Mr Ji and called the Sales name Xu Huauying.

Level: Crit

2.Xie Wenyu, Sales of Chengdu,who wanted to know whether the 8931GT ethernet device could take the place of 8931MT which had been stopped producting in PR6310 (SE7320SP2).I’ve told Du Xuewen how to test this device,and he did well as a frashman. Xie said the customer want to install WinXP in PR6310,but as the result,it would be blue screen during installing WinXP.Finnally,we used Windows 2000 server.

Level: Info

3.Zhang Zhihao,customer from Shanghai,asked some question about RAID.I told him SE7520 could not rebuild RAID data automatically,for example, in RAID 1 mode,if one harddisk was halted during the system running, it did not effect.But after reboot ,it would check the harddisk’s status and said one harddisk halted and the RAID was offline. We had to changed another harddisk whose capability must be bigger than the original,if it is smalled,the Rebuild function is unavailability.and then Enter the RAID configuration interface to rebuile the RAID in order to keep our data safe.
I’ve promised to send some documents about RAID to him in few days.

Level: Info

3.Reflashing logo of 7501CW2,Nanjing Fenghuo.
The picture format must be 640x480x4bpp and smaller than 320KB.In the pure Dos environment,issue the following command: a: 3.bat Xeon480.bmp

Level: Closed

4.Liu Quan,sales of Shanghai.SE7520BD2 only supports RAID0 and RAID1.If we want to create RAID10,the following RAID card would be choosed:
Adaptec 2200s, Intel SRCU42L,SRCU42E,SRCU42X ,
Mega Raid LSI 320-0,320-1,320-2.

5.Yang,inside sales of Shenzhen.The capability of NvRam in the SE7520BD2 is 8MB.This mainboard does not support Chipkill memory technology.SE7501 chipset supports 533MHz/400MHz FSB and SE7520 chipset supports 800Mhz FSB.
PC2100 = DDR266, PC 2700 = DDR333
LVD = interface of SCSI

Next:
——————————————
1.SATA – SATA2 conversion TEST (about TWO weeks).
Shu Ling will tell us which type harddish would be used and tested.
PT6250,PR2310,and then add 3ware RAID card in one of former two products,so we are going to prepare for three tests.
Level: Notice

2.JR2-DDR2 platform TEST.(release at 11th,Oct)
Using intel 2400 platform.
Level: Crit

3.Top 10 of R&D Lab(every Friday)