Configuration Files and Module Parameters (man5)

C H A P T E R 31

This section describes configuration files and module parameters and includes the following sections:

Introduction

Module Options

31.1 Introduction

LNET network hardware and routing are now configured via module parameters. Parameters should be specified in the /etc/modprobe.conf file, for example:

alias lustre llite
options lnet networks=tcp0,elan0

The above option specifies that this node should use all the available TCP and Elan interfaces.

Module parameters are read when the module is first loaded. Type-specific LND modules (for instance, ksocklnd) are loaded automatically by the LNET module when LNET starts (typically upon modprobe ptlrpc).

Under Linux 2.6, LNET configuration parameters can be viewed under /sys/module/; generic and acceptor parameters under LNET, and LND-specific parameters under the name of the corresponding LND.

Under Linux 2.4, sysfs is not available, but the LND-specific parameters are accessible via equivalent paths under /proc.

Important: All old (pre v.1.4.6) Lustre configuration lines should be removed from the module configuration files and replaced with the following. Make sure that CONFIG_KMOD is set in your linux.config so LNET can load the following modules it needs. The basic module files are:

modprobe.conf (for Linux 2.6)

alias lustre llite
options lnet networks=tcp0,elan0

modules.conf (for Linux 2.4)

alias lustre llite
options lnet networks=tcp0,elan0

For the following parameters, default option settings are shown in parenthesis. Changes to parameters marked with a W affect running systems. (Unmarked parameters can only be set when LNET loads for the first time.) Changes to parameters marked with Wc only have effect when connections are established (existing connections are not affected by these changes.)

31.2 Module Options

With routed or other multi-network configurations, use ip2nets rather than networks, so all nodes can use the same configuration.

For a routed network, use the same “routes” configuration everywhere. Nodes specified as routers automatically enable forwarding and any routes that are not relevant to a particular node are ignored. Keep a common configuration to guarantee that all nodes have consistent routing tables.

A separate modprobe.conf.lnet included from modprobe.conf makes distributing the configuration much easier.

If you set config_on_load=1, LNET starts at modprobe time rather than waiting for Lustre to start. This ensures routers start working at module load time.

# lctl 
# lctl> net down

Remember the lctl ping {nid} command - it is a handy way to check your LNET configuration.

31.2.1 LNET Options

This section describes LNET options.

31.2.1.1 Network Topology

Network topology module parameters determine which networks a node should join, whether it should route between these networks, and how it communicates with non-local networks.

Here is a list of various networks and the supported software stacks:

Network	Software Stack
openib	OpenIB gen1/Mellanox Gold
iib	Silverstorm (Infinicon)
vib	Voltaire
o2ib	OpenIB gen2
cib	Cisco
mx	Myrinet MX
gm	Myrinet GM-2
elan	Quadrics QSNet

Note - Lustre ignores the loopback interface (lo0), but Lustre use any IP addresses aliased to the loopback (by default). When in doubt, explicitly specify networks.

ip2nets ("") is a string that lists globally-available networks, each with a set of IP address ranges. LNET determines the locally-available networks from this list by matching the IP address ranges with the local IPs of a node. The purpose of this option is to be able to use the same modules.conf file across a variety of nodes on different networks. The string has the following syntax.

<ip2nets> :== <net-match> [ <comment> ] { <net-sep> <net-match> }
<net-match> :== [ <w> ] <net-spec> <w> <ip-range> { <w> <ip-range> }
[ <w> ]
<net-spec> :== <network> [ "(" <interface-list> ")" ]
<network> :== <nettype> [ <number> ]
<nettype> :== "tcp" | "elan" | "openib" | ...
<iface-list> :== <interface> [ "," <iface-list> ]
<ip-range> :== <r-expr> "." <r-expr> "." <r-expr> "." <r-expr>
<r-expr> :== <number> | "*" | "[" <r-list> "]"
<r-list> :== <range> [ "," <r-list> ]
<range> :== <number> [ "-" <number> [ "/" <number> ] ]
<comment :== "#" { <non-net-sep-chars> }
<net-sep> :== ";" | "\n"
<w> :== <whitespace-chars> { <whitespace-chars> }

<net-spec> contains enough information to uniquely identify the network and load an appropriate LND. The LND determines the missing "address-within-network" part of the NID based on the interfaces it can use.

<iface-list> specifies which hardware interface the network can use. If omitted, all interfaces are used. LNDs that do not support the <iface-list> syntax cannot be configured to use particular interfaces and just use what is there. Only a single instance of these LNDs can exist on a node at any time, and <iface-list> must be omitted.

<net-match> entries are scanned in the order declared to see if one of the node's IP addresses matches one of the <ip-range> expressions. If there is a match, <net-spec> specifies the network to instantiate. Note that it is the first match for a particular network that counts. This can be used to simplify the match expression for the general case by placing it after the special cases. For example:

ip2nets="tcp(eth1,eth2) 134.32.1.[4-10/2]; tcp(eth1) *.*.*.*"

4 nodes on the 134.32.1.* network have 2 interfaces (134.32.1.{4,6,8,10}) but all the rest have 1.

ip2nets="vib 192.168.0.*; tcp(eth2) 192.168.0.[1,7,4,12]"

This describes an IB cluster on 192.168.0.*. Four of these nodes also have IP interfaces; these four could be used as routers.

Note that match-all expressions (For instance, *.*.*.*) effectively mask all other
<net-match> entries specified after them. They should be used with caution.

Here is a more complicated situation, the route parameter is explained below. We have:

Two TCP subnets

One Elan subnet

One machine set up as a router, with both TCP and Elan interfaces

IP over Elan configured, but only IP will be used to label the nodes.

options lnet ip2nets=”tcp	198.129.135.* 192.128.88.98; elan  198.128.88.98 198.129.135.3;” routes=”tcp  1022@elan			# Elan NID of router; 		elan  198.128.88.98@tcp # TCP NID of router  “

31.2.1.2 networks ("tcp")

This is an alternative to "ip2nets" which can be used to specify the networks to be instantiated explicitly. The syntax is a simple comma separated list of <net-spec>s (see above). The default is only used if neither “ip2nets” nor “networks” is specified.

31.2.1.3 routes (“”)

This is a string that lists networks and the NIDs of routers that forward to them.

It has the following syntax (<w> is one or more whitespace characters):

<routes> :== <route>{ ; <route> }
<route> :== [<net>[<w><hopcount>]<w><nid>{<w><nid>}

So a node on the network tcp1 that needs to go through a router to get to the Elan network:

options lnet networks=tcp1 routes="elan 1 192.168.2.2@tcp1”

The hopcount is used to help choose the best path between multiply-routed configurations.

A simple but powerful expansion syntax is provided, both for target networks and router NIDs as follows.

<expansion> :== "[" <entry> { "," <entry> } "]"
<entry> :== <numeric range> | <non-numeric item>
<numeric range> :== <number> [ "-" <number> [ "/" <number> ] ]

The expansion is a list enclosed in square brackets. Numeric items in the list may be a single number, a contiguous range of numbers, or a strided range of numbers. For example, routes="elan 192.168.1.[22-24]@tcp" says that network elan0 is adjacent (hopcount defaults to 1); and is accessible via 3 routers on the tcp0 network (192.168.1.22@tcp, 192.168.1.23@tcp and 192.168.1.24@tcp).

routes="[tcp,vib] 2 [8-14/2]@elan" says that 2 networks (tcp0 and vib0) are accessible through 4 routers (8@elan, 10@elan, 12@elan and 14@elan). The hopcount of 2 means that traffic to both these networks will be traversed 2 routers - first one of the routers specified in this entry, then one more.

Duplicate entries, entries that route to a local network, and entries that specify routers on a non-local network are ignored.

Equivalent entries are resolved in favor of the route with the shorter hopcount. The hopcount, if omitted, defaults to 1 (the remote network is adjacent).

It is an error to specify routes to the same destination with routers on different local networks.

If the target network string contains no expansions, then the hopcount defaults to 1 and may be omitted (that is, the remote network is adjacent). In practice, this is true for most multi-network configurations. It is an error to specify an inconsistent hop count for a given target network. This is why an explicit hopcount is required if the target network string specifies more than one network.

31.2.1.4 forwarding ("")

This is a string that can be set either to "enabled" or "disabled" for explicit control of whether this node should act as a router, forwarding communications between all local networks.

A standalone router can be started by simply starting LNET (“modprobe ptlrpc”) with appropriate network topology options.

Variable	Description
acceptor	The acceptor is a TCP/IP service that some LNDs use to establish communications. If a local network requires it and it has not been disabled, the acceptor listens on a single port for connection requests that it redirects to the appropriate local network. The acceptor is part of the LNET module and configured by the following options: secure - Accept connections only from reserved TCP ports (< 1023). all - Accept connections from any TCP port. NOTE: this is required for liblustre clients to allow connections on non-privileged ports. none - Do not run the acceptor.
accept_port (988)	Port number on which the acceptor should listen for connection requests. All nodes in a site configuration that require an acceptor must use the same port.
accept_backlog (127)	Maximum length that the queue of pending connections may grow to (see listen(2)).
accept_timeout (5, W)	Maximum time in seconds the acceptor is allowed to block while communicating with a peer.
accept_proto_version	Version of the acceptor protocol that should be used by outgoing connection requests. It defaults to the most recent acceptor protocol version, but it may be set to the previous version to allow the node to initiate connections with nodes that only understand that version of the acceptor protocol. The acceptor can, with some restrictions, handle either version (that is, it can accept connections from both 'old' and 'new' peers). For the current version of the acceptor protocol (version 1), the acceptor is compatible with old peers if it is only required by a single local network.

31.2.2 SOCKLND Kernel TCP/IP LND

The SOCKLND kernel TCP/IP LND (socklnd) is connection-based and uses the acceptor to establish communications via sockets with its peers.

It supports multiple instances and load balances dynamically over multiple interfaces. If no interfaces are specified by the ip2nets or networks module parameter, all non-loopback IP interfaces are used. The address-within-network is determined by the address of the first IP interface an instance of the socklnd encounters.

Consider a node on the “edge” of an InfiniBand network, with a low-bandwidth management Ethernet (eth0), IP over IB configured (ipoib0), and a pair of GigE NICs (eth1,eth2) providing off-cluster connectivity. This node should be configured with "networks=vib,tcp(eth1,eth2)” to ensure that the socklnd ignores the management Ethernet and IPoIB.

Variable	Description
timeout (50,W)	Time (in seconds) that communications may be stalled before the LND completes them with failure.
nconnds (4)	Sets the number of connection daemons.
min_reconnectms (1000,W)	Minimum connection retry interval (in milliseconds). After a failed connection attempt, this is the time that must elapse before the first retry. As connections attempts fail, this time is doubled on each successive retry up to a maximum of 'max_reconnectms'.
max_reconnectms (6000,W)	Maximum connection retry interval (in milliseconds).
eager_ack (0 on linux, 1 on darwin,W)	Boolean that determines whether the socklnd should attempt to flush sends on message boundaries.
typed_conns (1,Wc)	Boolean that determines whether the socklnd should use different sockets for different types of messages. When clear, all communication with a particular peer takes place on the same socket. Otherwise, separate sockets are used for bulk sends, bulk receives and everything else.
min_bulk (1024,W)	Determines when a message is considered "bulk".
tx_buffer_size, rx_buffer_size (8388608,Wc)	Socket buffer sizes. Setting this option to zero (0), allows the system to auto-tune buffer sizes. WARNING: Be very careful changing this value as improper sizing can harm performance.
nagle (0,Wc)	Boolean that determines if nagle should be enabled. It should never be set in production systems.
keepalive_idle (30,Wc)	Time (in seconds) that a socket can remain idle before a keepalive probe is sent. Setting this value to zero (0) disables keepalives.
keepalive_intvl (2,Wc)	Time (in seconds) to repeat unanswered keepalive probes. Setting this value to zero (0) disables keepalives.
keepalive_count (10,Wc)	Number of unanswered keepalive probes before pronouncing socket (hence peer) death.
enable_irq_affinity (0,Wc)	Boolean that determines whether to enable IRQ affinity. The default is zero (0). When set, socklnd attempts to maximize performance by handling device interrupts and data movement for particular (hardware) interfaces on particular CPUs. This option is not available on all platforms. This option requires an SMP system to exist and produces best performance with multiple NICs. Systems with multiple CPUs and a single NIC may see increase in the performance with this parameter disabled.
zc_min_frag (2048,W)	Determines the minimum message fragment that should be considered for zero-copy sends. Increasing it above the platform's PAGE_SIZE disables all zero copy sends. This option is not available on all platforms.

31.2.3 QSW LND

The QSW LND (qswlnd) is connection-less and, therefore, does not need the acceptor. It is limited to a single instance, which uses all Elan "rails" that are present and dynamically load balances over them.

The address-with-network is the node's Elan ID. A specific interface cannot be selected in the "networks" module parameter.

Variable	Description
tx_maxcontig (1024)	Integer that specifies the maximum message payload (in bytes) to copy into a pre-mapped transmit buffer
mtxmsgs (8)	Number of "normal" message descriptors for locally-initiated communications that may block for memory (callers block when this pool is exhausted).
nnblk_txmsg (512 with a 4K page size, 256 otherwise)	Number of "reserved" message descriptors for communications that may not block for memory. This pool must be sized large enough so it is never exhausted.
nrxmsg_small (256)	Number of "small" receive buffers to post (typically everything apart from bulk data).
ep_envelopes_small (2048)	Number of message envelopes to reserve for the "small" receive buffer queue. This determines a breakpoint in the number of concurrent senders. Below this number, communication attempts are queued, but above this number, the pre-allocated envelope queue will fill, causing senders to back off and retry. This can have the unfortunate side effect of starving arbitrary senders, who continually find the envelope queue is full when they retry. This parameter should therefore be increased if envelope queue overflow is suspected.
nrxmsg_large (64)	Number of "large" receive buffers to post (typically for routed bulk data).
ep_envelopes_large (256)	Number of message envelopes to reserve for the "large" receive buffer queue. For more information on message envelopes, see the ep_envelopes_small option (above).
optimized_puts (32768,W)	Smallest non-routed PUT that will be RDMA’d.
optimized_gets (1,W)	Smallest non-routed GET that will be RDMA’d.

31.2.4 RapidArray LND

The RapidArray LND (ralnd) is connection-based and uses the acceptor to establish connections with its peers. It is limited to a single instance, which uses all (both) RapidArray devices present. It load balances over them using the XOR of the source and destination NIDs to determine which device to use for communication.

The address-within-network is determined by the address of the single IP interface that may be specified by the "networks" module parameter. If this is omitted, then the first non-loopback IP interface that is up is used instead.

Variable	Description
n_connd (4)	Sets the number of connection daemons.
min_reconnect_interval (1,W)	Minimum connection retry interval (in seconds). After a failed connection attempt, this sets the time that must elapse before the first retry. As connections attempts fail, this time is doubled on each successive retry, up to a maximum of the max_reconnect_interval option.
max_reconnect_interval (60,W)	Maximum connection retry interval (in seconds).
timeout (30,W)	Time (in seconds) that communications may be stalled before the LND completes them with failure.
ntx (64)	Number of "normal" message descriptors for locally-initiated communications that may block for memory (callers block when this pool is exhausted).
ntx_nblk (256)	Number of "reserved" message descriptors for communications that may not block for memory. This pool must be sized large enough so it is never exhausted.
fma_cq_size (8192)	Number of entries in the RapidArray FMA completion queue to allocate. It should be increased if the ralnd starts to issue warnings that the FMA CQ has overflowed. This is only a performance issue.
max_immediate (2048,W)	Size (in bytes) of the smallest message that will be RDMA’d, rather than being included as immediate data in an FMA. All messages greater than 6912 bytes must be RDMA’d (FMA limit).

31.2.5 VIB LND

The VIB LND is connection-based, establishing reliable queue-pairs over InfiniBand with its peers. It does not use the acceptor. It is limited to a single instance, using a single HCA that can be specified via the "networks" module parameter. If this is omitted, it uses the first HCA in numerical order it can open. The address-within-network is determined by the IPoIB interface corresponding to the HCA used.

Variable	Description
service_number (0x11b9a2)	Fixed IB service number on which the LND listens for incoming connection requests. NOTE: All instances of the viblnd on the same network must have the same setting for this parameter.
arp_retries (3,W)	Number of times the LND will retry ARP while it establishes communications with a peer.
min_reconnect_interval (1,W)	Minimum connection retry interval (in seconds). After a failed connection attempt, this sets the time that must elapse before the first retry. As connections attempts fail, this time is doubled on each successive retry, up to a maximum of the max_reconnect_interval option.
max_reconnect_interval (60,W)	Maximum connection retry interval (in seconds).
timeout (50,W)	Time (in seconds) that communications may be stalled before the LND completes them with failure.
ntx (32)	Number of "normal" message descriptors for locally-initiated communications that may block for memory (callers block when this pool is exhausted).
ntx_nblk (256	Number of "reserved" message descriptors for communications that may not block for memory. This pool must be sized large enough so it is never exhausted.
concurrent_peers (1152)	Maximum number of queue pairs and, therefore, the maximum number of peers that the instance of the LND may communicate with.
hca_basename ("InfiniHost")	Used to construct HCA device names by appending the device number.
ipif_basename ("ipoib")	Used to construct IPoIB interface names by appending the same device number as is used to generate the HCA device name.
local_ack_timeout (0x12,Wc)	Used to construct IPoIB interface names by appending the same device number as is used to generate the HCA device name.
retry_cnt (7,Wc)	Low-level QP parameter. Only change it from the default value if so advised.
rnr_cnt (6,Wc)	Low-level QP parameter. Only change it from the default value if so advised.
rnr_nak_timer (0x10,Wc)	Low-level QP parameter. Only change it from the default value if so advised.
fmr_remaps (1000)	Controls how often FMR mappings may be reused before they must be unmapped. Only change it from the default value if so advised
cksum (0,W)	Boolean that determines if messages (NB not RDMAs) should be check-summed. This is a diagnostic feature that should not normally be enabled.

31.2.6 OpenIB LND

The OpenIB LND is connection-based and uses the acceptor to establish reliable queue-pairs over InfiniBand with its peers. It is limited to a single instance that uses only IB device '0'.

The address-within-network is determined by the address of the single IP interface that may be specified by the "networks" module parameter. If this is omitted, the first non-loopback IP interface that is up, is used instead. It uses the acceptor to establish connections with its peers.

Variable	Description
n_connd (4)	Sets the number of connection daemons. The default value is 4.
min_reconnect_interval (1,W)	Minimum connection retry interval (in seconds). After a failed connection attempt, this sets the time that must elapse before the first retry. As connections attempts fail, this time is doubled on each successive retry, up to a maximum of 'max_reconnect_interval'.
max_reconnect_interval (60,W)	Maximum connection retry interval (in seconds).
timeout (50,W)	Time (in seconds) that communications may be stalled before the LND completes them with failure.
ntx (64)	Number of "normal" message descriptors for locally-initiated communications that may block for memory (callers block when this pool is exhausted).
ntx_nblk (256)	Number of "reserved" message descriptors for communications that may not block for memory. This pool must be sized large enough so it is never exhausted.
concurrent_peers (1024)	Maximum number of queue pairs and, therefore, the maximum number of peers that the instance of the LND may communicate with.
cksum (0,W)	Boolean that determines whether messages (NB not RDMAs) should be check-summed. This is a diagnostic feature that should not normally be enabled.

31.2.7 Portals LND (Linux)

The Portals LND Linux (ptllnd) can be used as a interface layer to communicate with Sandia Portals networking devices. This version is intended to work on Cray XT3 Linux nodes that use Cray Portals as a network transport.

Message Buffers

When ptllnd starts up, it allocates and posts sufficient message buffers to allow all expected peers (set by concurrent_peers) to send one unsolicited message. The first message that a peer actually sends is a

(so-called) "HELLO" message, used to negotiate how much additional buffering to setup (typically 8 messages). If 10000 peers actually exist, then enough buffers are posted for 80000 messages.

The maximum message size is set by the max_msg_size module parameter (default value is 512). This parameter sets the bulk transfer breakpoint. Below this breakpoint, payload data is sent in the message itself. Above this breakpoint, a buffer descriptor is sent and the receiver gets the actual payload.

The buffer size is set by the rxb_npages module parameter (default value is 1). The default conservatively avoids allocation problems due to kernel memory fragmentation. However, increasing this value to 2 is probably not risky.

The ptllnd also keeps an additional rxb_nspare buffers (default value is 8) posted to account for full buffers being handled.

Assuming a 4K page size with 10000 peers, 1258 buffers can be expected to be posted at startup, increasing to a maximum of 10008 as peers that are actually connected. By doubling rxb_npages halving max_msg_size, this number can be reduced by a factor of 4.

ME/MD Queue Length

The ptllnd uses a single portal set by the portal module parameter (default value of 9) for both message and bulk buffers. Message buffers are always attached with PTL_INS_AFTER and match anything sent with "message" matchbits. Bulk buffers are always attached with PTL_INS_BEFORE and match only specific matchbits for that particular bulk transfer.

This scheme assumes that the majority of ME / MDs posted are for "message" buffers, and that the overhead of searching through the preceding "bulk" buffers is acceptable. Since the number of "bulk" buffers posted at any time is also dependent on the bulk transfer breakpoint set by max_msg_size, this seems like an issue worth measuring at scale.

TX Descriptors

The ptllnd has a pool of so-called "tx descriptors", which it uses not only for outgoing messages, but also to hold state for bulk transfers requested by incoming messages. This pool should scale with the total number of peers.

To enable the building of the Portals LND (ptllnd.ko) configure with this option:

./configure --with-portals=<path-to-portals-headers>

Variable	Description
ntx (256)	Total number of messaging descriptors.
concurrent_peers (1152)	Maximum number of concurrent peers. Peers that attempt to connect beyond the maximum are not allowed.
peer_hash_table_size (101)	Number of hash table slots for the peers. This number should scale with concurrent_peers. The size of the peer hash table is set by the module parameter peer_hash_table_size which defaults to a value of 101. This number should be prime to ensure the peer hash table is populated evenly. It is advisable to increase this value to 1001 for ~10000 peers.
cksum (0)	Set to non-zero to enable message (not RDMA) checksums for outgoing packets. Incoming packets are always check-summed if necessary, independent of this value.
timeout (50)	Amount of time (in seconds) that a request can linger in a peers-active queue before the peer is considered dead.
portal (9)	Portal ID to use for the ptllnd traffic.
rxb_npages (64 * #cpus)	Number of pages in an RX buffer.
credits (128)	Maximum total number of concurrent sends that are outstanding to a single peer at a given time.
peercredits (8)	Maximum number of concurrent sends that are outstanding to a single peer at a given time.
max_msg_size (512)	Maximum immediate message size. This MUST be the same on all nodes in a cluster. A peer that connects with a different max_msg_size value will be rejected.

31.2.8 Portals LND (Catamount)

The Portals LND Catamount (ptllnd) can be used as a interface layer to communicate with Sandia Portals networking devices. This version is intended to work on the Cray XT3 Catamount nodes using Cray Portals as a network transport.

To enable the building of the Portals LND configure with this option:

./configure --with-portals=<path-to-portals-headers>

The following PTLLND tunables are currently available:

Variable	Description
PTLLND_DEBUG (boolean, dflt 0)	Enables or disables debug features.
PTLLND_TX_HISTORY (int, dflt debug?1024:0)	Sets the size of the history buffer.
PTLLND_ABORT_ON_PROTOCOL_MISMATCH (boolean, dflt 1)	Calls abort action on connecting to a peer running a different version of the ptllnd protocol.
PTLLND_ABORT_ON_NAK (boolean, dflt 0)	Calls abort action when a peer sends a NAK. (Example: When it has timed out this node.)
PTLLND_DUMP_ON_NAK (boolean, dflt debug?1:0)	Dumps peer debug and the history on receiving a NAK.
PTLLND_WATCHDOG_INTERVAL (int, dflt 1)	Sets intervals to check some peers for timed out communications while the application blocks for communications to complete.
PTLLND_TIMEOUT (int, dflt 50)	The communications timeout (in seconds).
PTLLND_LONG_WAIT (int, dflt debug?5:PTLLND_TIMEOUT)	The time (in seconds) after which the ptllnd prints a warning if it blocks for a longer time during connection establishment, cleanup after an error, or cleanup during shutdown.

The following environment variables can be set to configure the PTLLND’s behavior.

Variable	Description
PTLLND_PORTAL (9)	The portal ID (PID) to use for the ptllnd traffic.
PTLLND_PID (9)	The virtual PID on which to contact servers.
PTLLND_PEERCREDITS (8)	The maximum number of concurrent sends that are outstanding to a single peer at any given instant.
PTLLND_MAX_MESSAGE_SIZE (512)	The maximum messages size. This MUST be the same on all nodes in a cluster.
PTLLND_MAX_MSGS_PER_BUFFER (64)	The number of messages in a receive buffer. Receive buffer will be allocated of size PTLLND_MAX_MSGS_PER_BUFFER times PTLLND_MAX_MESSAGE_SIZE.
PTLLND_MSG_SPARE (256)	Additional receive buffers posted to portals.
PTLLND_PEER_HASH_SIZE (101)	Number of hash table slots for the peers.
PTLLND_EQ_SIZE (1024)	Size of the Portals event queue (that is, maximum number of events in the queue).

31.2.9 MX LND

MXLND supports a number of load-time parameters using Linux's module parameter system. The following variables are available:

Variable	Description
n_waitd	Number of completion daemons.
max_peers	Maximum number of peers that may connect.
cksum	Enables small message (< 4 KB) checksums if set to a non-zero value.
ntx	Number of total tx message descriptors.
credits	Number of concurrent sends to a single peer.
board	Index value of the Myrinet board (NIC).
ep_id	MX endpoint ID.
polling	Use zero (0) to block (wait). A value > 0 will poll that many times before blocking.
hosts	IP-to-hostname resolution file.

Of the described variables, only hosts is required. It must be the absolute path to the MXLND hosts file.

For example:

options kmxlnd hosts=/etc/hosts.mxlnd

The file format for the hosts file is:

IP  HOST  BOARD   EP_ID

The values must be space and/or tab separated where:

IP is a valid IPv4 address

HOST is the name returned by `hostname` on that machine

BOARD is the index of the Myricom NIC (0 for the first card, etc.)

EP_ID is the MX endpoint ID

To obtain the optimal performance for your platform, you may want to vary the remaining options.

n_waitd (1) sets the number of threads that process completed MX requests (sends and receives).

max_peers (1024) tells MXLND the upper limit of machines that it will need to communicate with. This affects how many receives it will pre-post and each receive will use one page of memory. Ideally, on clients, this value will be equal to the total number of Lustre servers (MDS and OSS). On servers, it needs to equal the total number of machines in the storage system. cksum (0) turns on small message checksums. It can be used to aid in troubleshooting. MX also provides an optional checksumming feature which can check all messages (large and small). For details, see the MX README.

ntx (256) is the number of total sends in flight from this machine. In actuality, MXLND reserves half of them for connect messages so make this value twice as large as you want for the total number of sends in flight.

credits (8) is the number of in-flight messages for a specific peer. This is part of the flow-control system in Lustre. Increasing this value may improve performance but it requires more memory because each message requires at least one page.

board (0) is the index of the Myricom NIC. Hosts can have multiple Myricom NICs and this identifies which one MXLND should use. This value must match the board value in your MXLND hosts file for this host.

ep_id (3) is the MX endpoint ID. Each process that uses MX is required to have at least one MX endpoint to access the MX library and NIC. The ID is a simple index starting at zero (0). This value must match the endpoint ID value in your MXLND hosts file for this host.

polling (0) determines whether this host will poll or block for MX request completions. A value of 0 blocks and any positive value will poll that many times before blocking. Since polling increases CPU usage, we suggest that you set this to zero (0) on the client and experiment with different values for servers.