Setting Up the Hosts
After meeting the hardware and software requirements, it is necessary to prepare the backend and client machines for installation of the Weka system. This preparation of the hosts consists of the following steps:
NIC driver installation.
SR-IOV enablement (when needed).
Network configuration.
Network configuration verification.
Clock synchronization.
Numa balancing disablement.
Note: Some of the examples in this page contain version-specific information. Since software is updated frequently, the package versions available to you may differ to those presented here.
Note: The steps describing the installation of NIC drivers are provided as a courtesy. Refer to your NIC vendor documentation for the latest information and updates.
This section describes an OFED installation procedure that has proven to be successful. However, Mellanox supports a number of other installation methods, any of which can be used to install OFED. For more information about other installation procedures, refer to the Mellanox documentation.
The Mellanox OFED installation has a number of dependencies. The following example shows the installation of OFED dependencies in RHEL/CentOS 7.x using yum's [base] and [update] repositories, which are supported and preconfigured in RHEL and CentOS.
yum install perl libnl lsof tcl libxml2-python tk
This example assumes that the server was provisioned using the "Minimal installation" option and that it has access to yum repositories, either locally or over the Internet. This method can trigger updates to existing packages already installed on the server.
Alternatively, it is possible to install OFED dependencies without triggering updates to already-installed packages, as shown in the following example:
yum --disablerepo=* --enablerepo=base install perl libnl lsof tcl libxml2-python tk
Once the dependencies have been satisfied, it is possible to perform the OFED installation procedure.
The Mellanox OFED installation involves decompressing the distribution archive (which should be obtained from the Mellanox website) and running the installation script. Refer to the following to begin the installation:
# tar xf MLNX_OFED_LINUX-4.5-1.0.1.0-rhel7.6-x86_64.tgz# ls -lFdrwxr-xr-x 6 root root 4096 Nov 28 2018 MLNX_OFED_LINUX-4.5-1.0.1.0-rhel7.6-x86_64/-rw-r--r-- 1 root root 239624023 Dec 2 2018 MLNX_OFED_LINUX-4.5-1.0.1.0-rhel7.5-x86_64.tgz# cd MLNX_OFED_LINUX-4.5-1.0.1.0-rhel7.6-x86_64/# ./mlnxofedinstall/tmp/MLNX_OFED_LINUX.414403.logsGeneral log file: /tmp/MLNX_OFED_LINUX.414403.logs/general.logVerifying KMP rpms compatibility with target kernel...This program will install the MLNX_OFED_LINUX package on your machine.Note that all other Mellanox, OEM, OFED, RDMA or Distribution IB packages will be removed.Those packages are removed due to conflicts with MLNX_OFED_LINUX, do not reinstall them.Do you want to continue?[y/N]:y
On completion of the OFED installation, the NIC firmware may be updated to match the firmware requirements of the Mellanox OFED software. If an update was performed, reboot the server at the end of the installation for the new firmware to become effective. Otherwise, restart the driver by running the following:
# /etc/init.d/openibd restart
This concludes the Mellanox OFED installation procedure.
SR-IOV enablement is not required for hosts with Mellanox NICs (CX-4 or newer).
SR-IOV enablement is mandatory for hosts equipped with Intel NICs, or when working with client VMs where there is a need to expose VFs of a physical NIC to the virtual NICs.
NetworkManager is a dynamic network control and configuration daemon. It is the default network management tool in some operating systems such as RHEL 6 and 7.
The Weka system requires network management to be handled by Network Initscripts (also known as "ifcfg configuration files"). This method is a basic network interface start/stop framework that is part of the initscripts package, and is the method that the Weka system currently supports in Red Hat and it derivatives.
The following commands can be used to permanently disable NetworkManager:
To stop NetworkManager:
systemctl stop NetworkManager
To disable NetworkManager:
systemctl disable NetworkManager
To validate NetworkManager is disabled
# systemctl is-enabled NetworkManagerdisabled
The following example of ifcfg script is provided a reference for configuring the Ethernet interface.
/etc/sysconfig/network-scripts/ifcfg-enp24s0TYPE="Ethernet"PROXY_METHOD="none"BROWSER_ONLY="no"BOOTPROTO="none"DEFROUTE="no"IPV4_FAILURE_FATAL="no"IPV6INIT="no"IPV6_AUTOCONF="no"IPV6_DEFROUTE="no"IPV6_FAILURE_FATAL="no"IPV6_ADDR_GEN_MODE="stable-privacy"NAME="enp24s0"DEVICE="enp24s0"ONBOOT="yes"NM_CONTROLLED=noIPADDR=192.168.1.1NETMASK=255.255.0.0MTU=9000
MTU 9000 (jumbo frame) is recommended for best performance. Refer to your switch vendor documentation for jumbo frame configuration.
Bring the interface up using the following command:
# ifup enp24s0
InfiniBand network configuration normally includes Subnet Manager (SM), but the procedure involved is beyond the scope of this document. However, it is important to be aware of the specifics of your SM configuration, such as partitioning and MTU, because they can affect the configuration of the endpoint ports in Linux. For best performance, MTU of 4092 is recommended.
Refer to the following ifcfg script when the IB network only has the default partition, i.e., "no pkey":
/etc/sysconfig/network-scripts/ifcfg-ib1TYPE=InfinibandONBOOT=yesBOOTPROTO=staticSTARTMODE=autoUSERCTL=noNM_CONTROLLED=noDEVICE=ib1IPADDR=192.168.1.1NETMASK=255.255.0.0MTU=4092
Bring the interface up using the following command:
# ifup ib1
Verify that the “default partition” connection is up, with all the attributes set:
# ip a s ib14: ib1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 4092 qdisc mq state UP group default qlen 256link/infiniband 00:00:03:72:fe:80:00:00:00:00:00:00:24:8a:07:03:00:a8:09:48brd 00:ff:ff:ff:ff:12:40:1b:ff:ff:00:00:00:00:00:00:ff:ff:ff:ffinet 10.0.20.84/24 brd 10.0.20.255 scope global noprefixroute ib0valid_lft forever preferred_lft forever
On an InfiniBand network with a non-default partition number, p-key must be configured on the interface if the InfiniBand ports on your network are members of an InfiniBand partition other than the default (0x7FFF). The p-key should associate the port as a full member of the partition (full members are those where the p-key number with the most-significant bit (MSB) of the 16-bits is set to 1).
For Example: If the partition number is 0x2, the limited member p-key will equal the p-key itself, i.e.,0x2. The full member p-key will be calculated as the logical OR of 0x8000 and the p-key (0x2) and therefore will be equal to 0x8002.
Note: All InfiniBand ports communicating with the Weka cluster must be full members.
Two ifcfg scripts must be created for each pkey-ed IPoIB interface. To determine your own pkey-ed IPoIB interface configuration, refer to the following two examples where a pkey of 0x8002 is used:
/etc/sysconfig/network-scripts/ifcfg-ib1TYPE=InfinibandONBOOT=yesMTU=4092BOOTPROTO=staticSTARTMODE=autoUSERCTL=noNM_CONTROLLED=noDEVICE=ib1
/etc/sysconfig/network-scripts/ifcfg-ib1.8002TYPE=InfinibandBOOTPROTO=noneCONNECTED_MODE=yesDEVICE=ib1.8002IPV4_FAILURE_FATAL=yesIPV6INIT=noMTU=4092NAME=ib1.8002NM_CONTROLLED=noONBOOT=yesPHYSDEV=ib1PKEY_ID=2PKEY=yesBROADCAST=192.168.255.255NETMASK=255.255.0.0IPADDR=192.168.1.1
Bring the interface up using the following command:
# ifup ib1.8002
Verify the connection is up with all the non-default partition attributes set:
# ip a s ib1.80025: ib1.8002@ib0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 4092 qdisc mq state UP qlen 256link/infiniband 00:00:11:03:fe:80:00:00:00:00:00:00:24:8a:07:03:00:a8:09:48 brd 00:ff:ff:ff:ff:12:40:1b:80:02:00:00:00:00:00:00:ff:ff:ff:ffinet 192.168.1.1/16 brd 192.168.255.255 scope global noprefixroute ib1.8002valid_lft forever preferred_lft forever
Use a large size ICMP ping to check the basic TCP/IP connectivity between the interfaces of the hosts:
# ping -M do -s 8972 -c 3 192.168.1.2PING 192.168.1.2 (192.168.1.2) 8972(9000) bytes of data.8980 bytes from 192.168.1.2: icmp_seq=1 ttl=64 time=0.063 ms8980 bytes from 192.168.1.2: icmp_seq=2 ttl=64 time=0.087 ms8980 bytes from 192.168.1.2: icmp_seq=3 ttl=64 time=0.075 ms--- 192.168.2.0 ping statistics ---3 packets transmitted, 3 received, 0% packet loss, time 1999msrtt min/avg/max/mdev = 0.063/0.075/0.087/0.009 ms
The-M do flag prohibits packet fragmentation, which allows verification of correct MTU configuration between the two end points.
-s 8972 is the maximum ICMP packet size that can be transferred with MTU 9000, due to the overhead of ICMP and IP protocols.
As described in Weka Networking HA section, bonded interfaces are not supported and each NIC must have its own IP address.
To configure Dual Network (IB or ETH), you will need to properly configure the routing of the interfaces involved.
Example using CentOS:
Add the following lines at the end of /etc/sysctl.conf:
net.ipv4.conf.ib0.arp_announce =2net.ipv4.conf.ib1.arp_announce =2net.ipv4.conf.ib0.arp_filter =1net.ipv4.conf.ib1.arp_filter =1
This can be added per interface, as described above, or to all interfaces:
net.ipv4.conf.all.arp_filter = 1net.ipv4.conf.default.arp_filter = 1net.ipv4.conf.all.arp_announce = 2net.ipv4.conf.default.arp_announce = 2
Routing Tables
Append the following to /etc/iproute2/rt_tables:
100 weka1101 weka2
Assuming the interfaces are mlnx0 and mlnx1 and assuming that the network is 10.90.0.0/16 with IPs 10.90.0.1 and 10.90.1.1 and a default gw of 10.90.2.1, set the following routing rules:
/etc/sysconfig/network-scripts/route-mlnx0
10.90.0.0/16 dev mlnx0 src 10.90.0.1 table weka1default via 10.90.2.1 dev mlnx0 table weka1
/etc/sysconfig/network-scripts/route-mlnx1
10.90.0.0/16 dev mlnx1 src 10.90.1.1 table weka2default via 10.90.2.1 dev mlnx1 table weka2
/etc/sysconfig/network-scripts/rule-mlnx0
table weka1 from 10.90.0.1
/etc/sysconfig/network-scripts/rule-mlnx1
table weka2 from 10.90.1.1
Refer to this link to learn how to configure dual Ethernet network in RHEL
The synchronization of time on computers and networks is considered good practice and is vitally important for the stability of the Weka system. Proper timestamp alignment in packets and logs is very helpful for the efficient and quick resolution of issues.
Configure the time synchronization software on the backend and client machines according to the specific vendor instructions (see your OS documentation), prior to installing the Weka software.
The Weka system manages the NUMA balancing by itself and makes the best possible decisions. Therefore, we recommend disabling the NUMA balancing feature of the Linux kernel to avoid additional latencies on the operations.
To disable NUMA balancing, run the following command on the host:
echo 0 > /proc/sys/kernel/numa_balancing