Below is a complete recipe how to build stylish, efficient and quiet mining pods from scratch. It is my second try at building crypto coin mining systems, after using a few mining-rig-in-a-crate units. The procedure was fully tested and works including somewhat tricky steps of blacklisting nouveau driver, and enabling NVIDIA fan speed control.
The overall procedure in this post is this:
[ ] Setup hardware [ ] Configure BIOS install CentOS 7.4 [ ] Install CUDA 9.1 [ ] Install Software to enable control of GPU fan speed [ ] Install mining software (example: XMR and ETH) [ ] Group servers into pods
Using open frame rigs has its advantages, but it will keep you busy maintaining it, and it just looks bad. It was time to build something for the longer term:
I selected the HP Z620 Workstation, an about 7 years old product with excellent specifications, including a 90% efficient power supply, two PCI-e 3.0 slots, dual LGA2011 socket Intel Xeon processor support, incl. 8-core CPUs with 20Mb Cache (Xeon E5-2650), great ECC support, system diagnostics, and on and on. Most importantly, refurbished servers and CPUs can be obtained at excellent low prices (if you search well), probably a sweet-spot for our purposes. The design of these units is truly first-class and it is a pleasure using and maintaining them.
Used Z620s come in two flavors, differed by BIOS boot block date. Boot block dated 2011 cannot be upgraded to support newer generation Xeon CPUs, and the boot block cannot be upgraded. This does not prevent you from updating the system BIOS, but the range of supported processors will remain limited. We are fine with that. Other than that, Z620 come with and without second CPU board, and if you purchase a unit without second board installed and later want to upgrade, you will have to hunt for a second board including a special bracket that replaces the top PCI-e expansion cover.
According to the specifications, the graphics cards alone may not consume more than in total 300W., i.e. two GTX 1070 cards, or perhaps a single GTX 1080 Ti. You would have to step up to the Z820 (service guide) with the 1.1kW PSU option for more, but at an higher server cost. The Z820 has 3 PCIe slots x16 and can accommodate at least 2 GTX 1080 Ti’s using a Molex to 8-pin conversion cable for the fourth GPU power connector. There is space for a third GPU but I have not tried it out. The Z820 is expensive as a starter system but upgradability seems almost unlimited.
The BOM, as of December 2017:
[ ] Refurbished HP Z620, 2x CPU E5-2603 1.8GHz, 16GB RAM $320 [ ] Refurbished 2x Intel Xeon E5-2650 2.0GHz CPUs, $110 [ ] Netis wireless bridge (4 ports) $15 [ ] 2x Nvidia FE GTX 1070 $800 Total: $1,245
Internals of an HP Z620 with two E5-2650 CPUs and two NVIDIA GTX 1070 FE GPUs.
The first steps are as follows:
[ ] Run MemTest86 Pro v7.4 Free, testing all cores and memory [ ] Rearrange the cables on the bottom of the case to allow insertion of large GTX cards into the lower full-size PCI-e slot. [ ] Upgrade BIOS to the latest version (we use J61 v3.91) [ ] Upgrade CPUs as needed and re-test MemTest86 [ ] Enter BIOS configuration menu using <F10> at boot [ ] Disable power savings options, in particular Enable PCIe Performance Mode
To protect the unit from dust we cover the front of the servers with a dust filter material.
Here are some of the BIOS configuration settings I used initially: SATA mode= AHCI; UEFI Boot Sources= Disabled (Booting from UEFA DVD-ROM device did not seem to work completely well the first time); Secure Boot= Disabled; DEP, VTx, VTd= Enabled, but Intel TXT(LT)= Disabled. OS Management of ESD= Enabled. Factory Recovery Boot Support= Disabled. (<F1> provides some minimal explanation for each of the options.)
On a second try, I found that UEFI Boot Sources= Enabled, and booting the CentOS7.4 DVD ROM as a UEFI Boot source works too without issue. However, building my own customized DVD ROM supporting UEFI booting was not so straightfoward.
Ethernet network (with a router’s DHCP server) was connected to the main LAN port I planned to use. The CentOS installation was tested first an Nvidia NVS 310 graphics card, and the second time with NVIDIA GTX 1050 Ti card installed. Ethernet network (with a router’s DHCP server) was connected to the main LAN port I planned to use.
It is possible to use Replicated Setup to read or write BIOS settings to a USB stick. Edit the resulting file, and remove the options you do not want to change when you read the configuration on another system.
While most miners seem to use easy to setup Ubuntu 16, we took the time to get things working on our preferred OS, CentOS 7.4 (1108). This previous post describes how to customize and build your own installer media but you can also use the standard “Minimal” installer. The page listing mirrors for ISO installer images is here. Pick a server close to you, or pick one directly from this list, or find what you want starting here.
[ ] Download the Minimal ISO image for CentOS 7.4 (e.g. CentOS-7-x86_64-Minimal-1708.iso - Sept. 13, 2017.) [ ] Optionally confirm the SHA256 checksum of the image file bba314624956961a2ea31dd460cd860a77911c1e0a56e4820a12b9c5dad363f5, on MacOS: shasum -a 256 <iso-file-path> [ ] Burn the ISO to a DVD (it's 830.5 MB)
The first time using CentOS 7.4 (1408) or 7.3 (1611) resulted in “An unknown error has occurred” in “storage.py at line 839” that prevented installation. This did not happen using a CentOS 7.1 (1503) installer. Selecting on the Error window “More info…” and [Debug] and then <Alt><Tab> to get to root shell prompt, I tried to erase any traces of former RAID configuration data on the (refurbished) drive:
dmraid -r -E /dev/sda <y> reboot
Booting the installer again, the issue was resolved and installation of CentOS 7.4 could proceed. Set the installation destination (reclaim space, deleting all data, if needed), Ethernet interface (select the plugged in LAN port, turn ON (this will show IP address from the DHCP server), set host name (e.g. my_host_name.localdomain), and optionally Configure static IP (IPv4 – Manual; Add address, net mask, gateway), timezone, keyboard (I added French, Canadian Multilingual), etc. Then [Begin Installation] after which you can take your time to set Root password and create your user account (and make yourself user administrator by checking the box).
CentOS7.4 by default seems to require you to accept an empty EULA on the terminal once, but this does not impact any other operation of the system.
Boot the system, login on the terminal, and note the IP address (“inet”) of the LAN interface:
Next we use our terminal from MacOS and ssh to login and configure the system remotely. (Delete the line with the IP address of the new system from the ~/.ssh/known_hosts file if the IP address was used before with another server and LAN interface.)
CUDA 9.1 Installation
CUDA 9.1 (9.1.85-1) with driver version 387.26 (which does not seem to display consumed power anymore though for 1050Ti cards) is available and it seems to be working with the other software we plan to use so we can stop using CUDA 8.0 (we used 8.0.61-1 before). For additional reference, NVIDIAs instructions for CentOS/RHEL are here. The archive with all version is here.
(You can view available software version with sudo yum –showduplicates list <package_name> and specifically install a particular version by adding version to package name.)
[ ] Download RPM(local) version of CUDA9.1 from NVIDIA here to your remote PC [Linux][x86_64][CentOS][rpm (local)][Download] [ ] Copy the file from your remote PC to the server, typing on the remote PC (MacOS) scp cuda-repo-rhel7-9-1-local-9.1.85-1.x86_64.rpm <userid>@<ip_address>:/home/<userid>/ [ ] Login to your server (assuming your user id on remote and server are the same) ssh <ip_address> [ ] Install rpm sudo rpm -i cuda-repo-rhel7-9-1-local-9.1.85-1.x86_64.rpm [ ] Install epel-release repository (tested version 7-9) and dkms (version 2.4.0-1) sudo yum install epel-release -y sudo yum install dkms -y [ ] Install kernel headers and dev packages (our kernel was 3.10.0-693.el7.x86_64) sudo yum install kernel-devel-$(uname -r) kernel-headers-$(uname -r) -y [ ] Install cuda-9.1.85-1 (or just cuda if you only have one version available) sudo yum install cuda-9.1.85-1 -y
Backlisting Open Source drivers
Next we will have to blacklist the open source Nouveau driver.
[ ] Edit /etc/modprobe.d/blacklist-nouveau.conf sudo vi /etc/modprobe.d/blacklist-nouveau.conf [ ] Add to the file so that it includes the following blacklist nouveau blacklist lbm-nouveau options nouveau modeset=0 alias nouveau off alias lbm-nouveau off [ ] Save and exit [ ] Backup the current initramfs sudo mv /boot/initramfs-$(uname -r).img /boot/initramfs-$(uname -r)-nouveau.img [ ] Create the new initramfs (in Ubuntu it is sudo update-initramfs -u) sudo dracut /boot/initramfs-$(uname -r).img $(uname -r) [ ] Reboot sudo reboot [ ] Login and try if this lists the status of the GPU(s) nvidia-smi
Install screen (version 4.1.0-0.23) to run processes in the background and connect to them easily at any time.
sudo yum install screen -y
Create a screen running a process with screen -dmS <screen_name> <process_command>
Enter the screen to connect to using screen -r <screen_name> . And disconnect from the screen with <Ctrl><a><d> which leaves the process running in the background even when you log off!
To control the fan speed and certain other parameters of the NVIDIA cards you will need a running Xorg.
sudo yum -y groupinstall "X Window System"
These are the software versions installed:
glx-utils x86_64 8.2.0-3 initial-setup-gui x86_64 0.3.9.40-1 mesa-dri-drivers x86_64 17.0.1-6.20170307 plymouth-system-theme x86_64 0.8.9-0.28.20140113 spice-vdagent x86_64 0.14.0-14 xorg-x11-drivers x86_64 7.7-6 xorg-x11-server-Xorg x86_64 1.19.3-11 xorg-x11-utils x86_64 7.5-22 xorg-x11-xinit x86_64 1.3.4-1 xvattr x86_64 1.3-27
Enable GPU Control
To enable control of certain attributes, the parameter coolbits needs to be set to 7 for fan speed control only enabled, or 31 for everything enabled.
[ ] Enable "coolbits" to get access to certain control attributes sudo nvidia-xconfig --enable-all-gpus --allow-empty-initial-configuration --cool-bits=31 [ ] Reboot sudo reboot now
After adding a GPU later, you need to execute this again to add entries for the new card.
CUDA Miner Installation
A CUDA miner for Monero (XMR) is built from source as follows:
NOTE: As of begin April 2018 the Cryptonight algorithm that Monero uses has changed to shut out ASIC miners. The following procedure works with git reset –hard ef87d28 to check out version 2.3.0 or git reset –hard 26a5d65 for version 2.4.3, both with monero7 / modified algorithm. Select to mine “monero7”.
[ ] Install scl (software collections) (centos-release-scl 0:2-2) sudo yum install centos-release-scl -y [ ] Install git (version 0:18.104.22.168-12) sudo yum install git -y [ ] Install cmake3 (0:3.6.3-1), devtoolset-4-gcc* (0:5.3.1-6.1), hwloc-devel (1.11.2-2), libmicrohttpd-devel (0.9.33-2), openssl-devel (1:1.0.2k-8) sudo yum install cmake3 devtoolset-4-gcc* hwloc-devel libmicrohttpd-devel openssl-devel make -y [ ] Enter SCL bash shell scl enable devtoolset-4 bash [ ] Clone the repository and enter the top directory git clone https://github.com/fireice-uk/xmr-stak.git cd xmr-stak [ ] pick commit 7584ea8 version 2.2.0 (or skip this for newest but untested) git reset --hard 7584ea8 [ ] Optionally adjust donation level to developer vi xmrstak/donate-level.hpp [ ] Create build directory and change directory mkdir build cd build [ ] Build for CUDA cmake3 -DOpenCL_ENABLE=OFF .. make install [ ] Exit SCL bash exit
For CPU mining increase memlock limits, and enable huge pages, then reboot:
[ ] Append to file /etc/security/limits.conf sudo vi /etc/security/limits.conf * soft memlock 262144 * hard memlock 262144 [ ] Append to file /etc/sysctl.conf sudo vi /etc/sysctl.conf vm.nr_hugepages=128 [ ] Reboot sudo reboot now
Build a CUDA miner for Ethereum from source like this:
[ ] Change directory to top cd ~ [ ] Clone repo git clone https://github.com/ethereum-mining/ethminer.git [ ] Select a specific version (or skip this for newest but untested) cd ethminer [ ] update submodules git submodule update --init --recursive [ ] select commit (tested 0.14.0rc4 (3eaf0ec), 0.13.0.dev0 (5bd2b34)) git reset --hard 3eaf0ec [ ] Build for CUDA mkdir build; cd build cmake3 -DETHASHCL=OFF -DETHASHCUDA=ON .. sudo make install [ ] Confirm the location of the miner here: /usr/local/bin/ethminer ls /usr/local/bin/ethminer [ ] Or use the copy in ethminer/build/ethminer/
Refer to the Github pages or mining pool help sections for information how to configure and run the miners.
Standard Mining Pods
To protect servers from dust, buy some filter media and cover the front of each. I bought a 90 ft. roll of Air Handler Filter Media MERV 8 material of a width similar to that of my servers at reasonable cost around $25. Cut a bit longer than the server’s height, tuck one edge under the front feet and pull the filter media tight through behind the front handle of the server. Very cheap and white side out does not look too bad.
You can group 2-3 servers together into a pod, depending on the GPUs and their power consumption. Connect the servers to a wireless bridge, a quality surge protector, and an optional Watt meter and power switch, and place them on an electrical circuit separate from any other pods.
Caution: You are responsible for avoiding overloading your specific electrical system (or worse), and for safety in general; do not do this yourself if you do not have the necessary experience.
Mining pod consisting of three HP Z620s with various GPU configurations, looking good.
updated: 20180128; 20180207; 20180330; 20180401; 20180411; 20180504; 20180531
photo: ideanist M