12. DPDK Release 18.05
12.1. New Features
Reworked memory subsystem.
Memory subsystem has been reworked to support new functionality.
On Linux, support for reserving/unreserving hugepage memory at runtime has been added, so applications no longer need to pre-reserve memory at startup. Due to reorganized internal workings of memory subsystem, any memory allocated through
rte_malloc()
orrte_memzone_reserve()
is no longer guaranteed to be IOVA-contiguous.This functionality has introduced the following changes:
rte_eal_get_physmem_layout()
was removed.A new flag for memzone reservation (
RTE_MEMZONE_IOVA_CONTIG
) was added to ensure reserved memory will be IOVA-contiguous, for use with device drivers and other cases requiring such memory.New callbacks for memory allocation/deallocation events, allowing users (or drivers) to be notified of new memory being allocated or deallocated
New callbacks for validating memory allocations above a specified limit, allowing user to permit or deny memory allocations.
A new command-line switch
--legacy-mem
to enable EAL behavior similar to how older versions of DPDK worked (memory segments that are IOVA-contiguous, but hugepages are reserved at startup only, and can never be released).A new command-line switch
--single-file-segments
to put all memory segments within a segment list in a single file.A set of convenience function calls to look up and iterate over allocated memory segments.
-m
and--socket-mem
command-line arguments now carry an additional meaning and mark pre-reserved hugepages as “unfree-able”, thereby acting as a mechanism guaranteeing minimum availability of hugepage memory to the application.
Reserving/unreserving memory at runtime is not currently supported on FreeBSD.
Added bucket mempool driver.
Added a bucket mempool driver which provides a way to allocate contiguous block of objects. The number of objects in the block depends on how many objects fit in the
RTE_DRIVER_MEMPOOL_BUCKET_SIZE_KB
memory chunk which is a build time option. The number may be obtained usingrte_mempool_ops_get_info()
API. Contiguous blocks may be allocated usingrte_mempool_get_contig_blocks()
API.Added support for port representors.
Added DPDK port representors (also known as “VF representors” in the specific context of VFs), which are to DPDK what the Ethernet switch device driver model (switchdev) is to Linux, and which can be thought as a software “patch panel” front-end for applications. DPDK port representors are implemented as additional virtual Ethernet device (ethdev) instances, spawned on an as-needed basis through configuration parameters passed to the driver of the underlying device using devargs.
Added support for VXLAN and NVGRE tunnel endpoint.
New actions types have been added to support encapsulation and decapsulation operations for a tunnel endpoint. The new action types are
RTE_FLOW_ACTION_TYPE_[VXLAN/NVGRE]_ENCAP
,RTE_FLOW_ACTION_TYPE_[VXLAN/NVGRE]_DECAP
,RTE_FLOW_ACTION_TYPE_JUMP
. A new item typeRTE_FLOW_ACTION_TYPE_MARK
has been added to match a flow against a previously marked flow. A shared counter has also been introduced to the flow API to count a group of flows.Added PMD-recommended Tx and Rx parameters.
Applications can now query drivers for device-tuned values of ring sizes, burst sizes, and number of queues.
Added RSS hash and key update to CXGBE PMD.
Added support for updating the RSS hash and key to the CXGBE PMD.
Added CXGBE VF PMD.
CXGBE VF Poll Mode Driver has been added to run DPDK over Chelsio T5/T6 NIC VF instances.
Updated mlx5 driver.
Updated the mlx5 driver including the following changes:
Introduced Multi-packet Rx to enable 100Gb/sec with 64B frames.
Support for being run by non-root users given a reduced set of capabilities
CAP_NET_ADMIN
,CAP_NET_RAW
andCAP_IPC_LOCK
.Support for TSO and checksum for generic UDP and IP tunnels.
Support for inner checksum and RSS for GRE, VXLAN-GPE, MPLSoGRE and MPLSoUDP tunnels.
Accommodate the new memory hotplug model.
Support for non virtually contiguous mempools.
Support for MAC adding along with allmulti and promiscuous modes from VF.
Support for Mellanox BlueField SoC device.
Support for PMD defaults for queue number and depth to improve the out of the box performance.
Updated mlx4 driver.
Updated the mlx4 driver including the following changes:
Support for to being run by non-root users given a reduced set of capabilities
CAP_NET_ADMIN
,CAP_NET_RAW
andCAP_IPC_LOCK
.Supported CRC strip toggling.
Accommodate the new memory hotplug model.
Support non virtually contiguous mempools.
Dropped support for Mellanox OFED 4.2.
Updated Solarflare network PMD.
Updated the sfc_efx driver including the following changes:
Added support for Solarflare XtremeScale X2xxx family adapters.
Added support for NVGRE, VXLAN and GENEVE filters in flow API.
Added support for DROP action in flow API.
Added support for equal stride super-buffer Rx mode (X2xxx only).
Added support for MARK and FLAG actions in flow API (X2xxx only).
Added Ethernet poll mode driver for AMD XGBE devices.
Added the new
axgbe
ethernet poll mode driver for AMD XGBE devices. See the AXGBE Poll Mode Driver nic driver guide for more details on this new driver.Updated szedata2 PMD.
Added support for new NFB-200G2QL card. A new API was introduced in the libsze2 library which the szedata2 PMD depends on, thus the new version of the library was needed. New versions of the packages are available and the minimum required version is 4.4.1.
Added support for Broadcom NetXtreme-S (BCM58800) family of controllers (aka Stingray).
Added support for the Broadcom NetXtreme-S (BCM58800) family of controllers (aka Stingray). The BCM58800 devices feature a NetXtreme E-Series advanced network controller, a high-performance ARM CPU block, PCI Express (PCIe) Gen3 interfaces, key accelerators for compute offload and a high-speed memory subsystem including L3 cache and DDR4 interfaces, all interconnected by a coherent Network-on-chip (NOC) fabric.
The ARM CPU subsystem features eight ARMv8 Cortex-A72 CPUs at 3.0 GHz, arranged in a multi-cluster configuration.
Added vDPA in vhost-user lib.
Added support for selective datapath in the vhost-user lib. vDPA stands for vhost Data Path Acceleration. It supports virtio ring compatible devices to serve the virtio driver directly to enable datapath acceleration.
Added IFCVF vDPA driver.
Added IFCVF vDPA driver to support Intel FPGA 100G VF devices. IFCVF works as a HW vhost data path accelerator, it supports live migration and is compatible with virtio 0.95 and 1.0. This driver registers the ifcvf vDPA driver to vhost lib, when virtio connects. With the help of the registered vDPA driver the assigned VF gets configured to Rx/Tx directly to VM’s virtio vrings.
Added support for vhost dequeue interrupt mode.
Added support for vhost dequeue interrupt mode to release CPUs to others when there is no data to transmit. Applications can register an epoll event file descriptor to associate Rx queues with interrupt vectors.
Added support for virtio-user server mode.
In a container environment if the vhost-user backend restarts, there’s no way for it to reconnect to virtio-user. To address this, support for server mode has been added. In this mode the socket file is created by virtio-user, which the backend connects to. This means that if the backend restarts, it can reconnect to virtio-user and continue communications.
Added crypto workload support to vhost library.
New APIs have been introduced in the vhost library to enable virtio crypto support including session creation/deletion handling and translating virtio-crypto requests into DPDK crypto operations. A sample application has also been introduced.
Added virtio crypto PMD.
Added a new Poll Mode Driver for virtio crypto devices, which provides AES-CBC ciphering and AES-CBC with HMAC-SHA1 algorithm-chaining. See the Virtio Crypto Poll Mode Driver crypto driver guide for more details on this new driver.
Added AMD CCP Crypto PMD.
Added the new
ccp
crypto driver for AMD CCP devices. See the AMD CCP Poll Mode Driver crypto driver guide for more details on this new driver.Updated AESNI MB PMD.
The AESNI MB PMD has been updated with additional support for:
AES-CMAC (128-bit key).
Added the Compressdev Library, a generic compression service library.
Added the Compressdev library which provides an API for offload of compression and decompression operations to hardware or software accelerator devices.
Added a new compression poll mode driver using Intels ISA-L.
Added the new
ISA-L
compression driver, for compression and decompression operations in software. See the ISA-L Compression Poll Mode Driver compression driver guide for details on this new driver.Added the Event Timer Adapter Library.
The Event Timer Adapter Library extends the event-based model by introducing APIs that allow applications to arm/cancel event timers that generate timer expiry events. This new type of event is scheduled by an event device along with existing types of events.
Added OcteonTx TIM Driver (Event timer adapter).
The OcteonTx Timer block enables software to schedule events for a future time, it is exposed to an application via the Event timer adapter library.
See the OCTEON TX SSOVF Eventdev Driver guide for more details
Added Event Crypto Adapter Library.
Added the Event Crypto Adapter Library. This library extends the event-based model by introducing APIs that allow applications to enqueue/dequeue crypto operations to/from cryptodev as events scheduled by an event device.
Added Ifpga Bus, a generic Intel FPGA Bus library.
Added the Ifpga Bus library which provides support for integrating any Intel FPGA device with the DPDK framework. It provides Intel FPGA Partial Bit Stream AFU (Accelerated Function Unit) scan and drivers probe.
Added IFPGA (Intel FPGA) Rawdev Driver.
Added a new Rawdev driver called IFPGA (Intel FPGA) Rawdev Driver, which cooperates with OPAE (Open Programmable Acceleration Engine) shared code to provide common FPGA management ops for FPGA operation.
See the IFPGA Rawdev Driver programmer’s guide for more details.
Added DPAA2 QDMA Driver (in rawdev).
The DPAA2 QDMA is an implementation of the rawdev API, that provide a means of initiating a DMA transaction from CPU. The initiated DMA is performed without the CPU being involved in the actual DMA transaction.
See the NXP DPAA2 QDMA Driver guide for more details.
Added DPAA2 Command Interface Driver (in rawdev).
The DPAA2 CMDIF is an implementation of the rawdev API, that provides communication between the GPP and NXP’s QorIQ based AIOP Block (Firmware). Advanced IO Processor i.e. AIOP are clusters of programmable RISC engines optimized for flexible networking and I/O operations. The communication between GPP and AIOP is achieved via using DPCI devices exposed by MC for GPP <–> AIOP interaction.
See the NXP DPAA2 CMDIF Driver guide for more details.
Added device event monitor framework.
Added a general device event monitor framework to EAL, for device dynamic management to facilitate device hotplug awareness and associated actions. The list of new APIs is:
rte_dev_event_monitor_start
andrte_dev_event_monitor_stop
for the event monitor enabling and disabling.rte_dev_event_callback_register
andrte_dev_event_callback_unregister
for registering and un-registering user callbacks.
Linux uevent is supported as a backend of this device event notification framework.
Added support for procinfo and pdump on eth vdev.
For ethernet virtual devices (like TAP, PCAP, etc.), with this feature, we can get stats/xstats on shared memory from a secondary process, and also pdump packets on those virtual devices.
Enhancements to the Packet Framework Library.
Design and development of new API functions for Packet Framework library that implement a common set of actions such as traffic metering, packet encapsulation, network address translation, TTL update, etc., for pipeline table and input ports to speed up application development. The API functions includes creating action profiles, registering actions to the profiles, instantiating action profiles for pipeline table and input ports, etc.
Added the BPF Library.
The BPF Library provides the ability to load and execute Enhanced Berkeley Packet Filters (eBPF) within user-space DPDK applications. It also introduces a basic framework to load/unload BPF-based filters on Eth devices (right now only via SW RX/TX callbacks). It also adds a dependency on libelf.
12.2. API Changes
service cores: No longer marked as experimental.
The service cores functions are no longer marked as experimental, and have become part of the normal DPDK API and ABI. Any future ABI changes will be announced at least one release before the ABI change is made. There are no ABI breaking changes planned.
eal: The
rte_lcore_has_role()
return value changed.This function now returns true or false, respectively, rather than 0 or < 0 for success or failure. It makes use of the function more intuitive.
mempool: The capability flags and related functions have been removed.
Flags
MEMPOOL_F_CAPA_PHYS_CONTIG
andMEMPOOL_F_CAPA_BLK_ALIGNED_OBJECTS
were used by octeontx mempool driver to customize generic mempool library behavior. Now the new driver callbackscalc_mem_size
andpopulate
may be used to achieve it without specific knowledge in the generic code.mempool: The following xmem functions have been deprecated:
rte_mempool_xmem_create
rte_mempool_xmem_size
rte_mempool_xmem_usage
rte_mempool_populate_iova_tab
mbuf: The control mbuf API has been removed in v18.05. The impacted functions and macros are:
rte_ctrlmbuf_init()
rte_ctrlmbuf_alloc()
rte_ctrlmbuf_free()
rte_ctrlmbuf_data()
rte_ctrlmbuf_len()
rte_is_ctrlmbuf()
CTRL_MBUF_FLAG
The packet mbuf API should be used as a replacement.
meter: API updated to accommodate configuration profiles.
The meter API has been changed to support meter configuration profiles. The configuration profile represents the set of configuration parameters for a given meter object, such as the rates and sizes for the token buckets. These configuration parameters were previously part of the meter object internal data structure. The separation of the configuration parameters from the meter object data structure results in reducing its memory footprint which helps in better cache utilization when a large number of meter objects are used.
ethdev: The function
rte_eth_dev_count()
, often mis-used to iterate over ports, is deprecated and replaced byrte_eth_dev_count_avail()
. There is also a new functionrte_eth_dev_count_total()
to get the total number of allocated ports, available or not. The hotplug-proof applications should useRTE_ETH_FOREACH_DEV
orRTE_ETH_FOREACH_DEV_OWNED_BY
as port iterators.ethdev: In struct
struct rte_eth_dev_info
, fieldrte_pci_device *pci_dev
has been replaced with fieldstruct rte_device *device
.ethdev: Changes to the semantics of
rte_eth_dev_configure()
parameters.If both the
nb_rx_q
andnb_tx_q
parameters are zero,rte_eth_dev_configure()
will now use PMD-recommended queue sizes, or if recommendations are not provided by the PMD the function will use ethdev fall-back values. Previously setting both of the parameters to zero would have resulted in-EINVAL
being returned.ethdev: Changes to the semantics of
rte_eth_rx_queue_setup()
parameters.If the
nb_rx_desc
parameter is zero,rte_eth_rx_queue_setup
will now use the PMD-recommended Rx ring size, or in the case where the PMD does not provide a recommendation, will use an ethdev-provided fall-back value. Previously, settingnb_rx_desc
to zero would have resulted in an error.ethdev: Changes to the semantics of
rte_eth_tx_queue_setup()
parameters.If the
nb_tx_desc
parameter is zero,rte_eth_tx_queue_setup
will now use the PMD-recommended Tx ring size, or in the case where the PMD does not provide a recommendation, will use an ethdev-provided fall-back value. Previously, settingnb_tx_desc
to zero would have resulted in an error.ethdev: Several changes were made to the flow API.
The unused DUP action was removed.
Actions semantics in flow rules: list order now matters (“first to last” instead of “all simultaneously”), repeated actions are now all performed, and they do not individually have (non-)terminating properties anymore.
Flow rules are now always terminating unless a
PASSTHRU
action is present.C99-style flexible arrays were replaced with standard pointers in RSS action and in RAW pattern item structures due to compatibility issues.
The RSS action was modified to not rely on external
struct rte_eth_rss_conf
anymore to instead expose its own and more appropriately named configuration fields directly (rss_conf->rss_key
=>key
,rss_conf->rss_key_len
=>key_len
,rss_conf->rss_hf
=>types
,num
=>queue_num
), and the addition of missing RSS parameters (func
for RSS hash function to apply andlevel
for the encapsulation level).The VLAN pattern item (
struct rte_flow_item_vlan
) was modified to include inner EtherType instead of outer TPID. Its default mask was also modified to cover the VID part (lower 12 bits) of TCI only.A new transfer attribute was added to
struct rte_flow_attr
in order to clarify the behavior of some pattern items.PF and VF pattern items are now only accepted by PMDs that implement them (bnxt and i40e) when the transfer attribute is also present, for consistency.
Pattern item PORT was renamed PHY_PORT to avoid confusion with DPDK port IDs.
An action counterpart to the PHY_PORT pattern item was added in order to redirect matching traffic to a specific physical port.
PORT_ID pattern item and actions were added to match and target DPDK port IDs at a higher level than PHY_PORT.
RTE_FLOW_ACTION_TYPE_[VXLAN/NVGRE]_ENCAP
action items were added to support tunnel encapsulation operation for VXLAN and NVGRE type tunnel endpoint.RTE_FLOW_ACTION_TYPE_[VXLAN/NVGRE]_DECAP
action items were added to support tunnel decapsulation operation for VXLAN and NVGRE type tunnel endpoint.RTE_FLOW_ACTION_TYPE_JUMP
action item was added to support a matched flow to be redirected to the specific group.RTE_FLOW_ACTION_TYPE_MARK
item type has been added to match a flow against a previously marked flow.
ethdev: Change flow APIs regarding count action:
rte_flow_create()
API count action now requires thestruct rte_flow_action_count
.rte_flow_query()
API parameter changed from action type to action structure.
ethdev: Changes to offload API
A pure per-port offloading isn’t requested to be repeated in [rt]x_conf->offloads to
rte_eth_[rt]x_queue_setup()
. Now any offloading enabled inrte_eth_dev_configure()
can’t be disabled byrte_eth_[rt]x_queue_setup()
. Any new added offloading which has not been enabled inrte_eth_dev_configure()
and is requested to be enabled inrte_eth_[rt]x_queue_setup()
must be per-queue type, or otherwise trigger an error log.ethdev: Runtime queue setup
rte_eth_rx_queue_setup
andrte_eth_tx_queue_setup
can be called afterrte_eth_dev_start
if the device supports runtime queue setup. The device driver can expose this capability throughrte_eth_dev_info_get
. A Rx or Tx queue set up at runtime need to be started explicitly byrte_eth_dev_rx_queue_start
orrte_eth_dev_tx_queue_start
.
12.3. ABI Changes
ring: The alignment constraints on the ring structure has been relaxed to one cache line instead of two, and an empty cache line padding is added between the producer and consumer structures. The size of the structure and the offset of the fields remains the same on platforms with 64B cache line, but changes on other platforms.
mempool: Some ops have changed.
A new callback
calc_mem_size
has been added torte_mempool_ops
to allow customization of the required memory size calculation. A new callbackpopulate
has been added torte_mempool_ops
to allow customized object population. Callbackget_capabilities
has been removed fromrte_mempool_ops
since its features are covered bycalc_mem_size
andpopulate
callbacks. Callbackregister_memory_area
has been removed fromrte_mempool_ops
since the new callbackpopulate
may be used instead of it.ethdev: Additional fields in rte_eth_dev_info.
The
rte_eth_dev_info
structure has had two extra entries appended to the end of it:default_rxportconf
anddefault_txportconf
. Each of these in turn arerte_eth_dev_portconf
structures containing three fields of typeuint16_t
:burst_size
,ring_size
, andnb_queues
. These are parameter values recommended for use by the PMD.ethdev: ABI for all flow API functions was updated.
This includes functions
rte_flow_copy
,rte_flow_create
,rte_flow_destroy
,rte_flow_error_set
,rte_flow_flush
,rte_flow_isolate
,rte_flow_query
andrte_flow_validate
, due to changes in error type definitions (enum rte_flow_error_type
), removal of the unused DUP action (enum rte_flow_action_type
), modified behavior for flow rule actions (see API changes), removal of C99 flexible array from RAW pattern item (struct rte_flow_item_raw
), complete rework of the RSS action definition (struct rte_flow_action_rss
), sanity fix in the VLAN pattern item (struct rte_flow_item_vlan
) and new transfer attribute (struct rte_flow_attr
).bbdev: New parameter added to rte_bbdev_op_cap_turbo_dec.
A new parameter
max_llr_modulus
has been added torte_bbdev_op_cap_turbo_dec
structure to specify maximal LLR (likelihood ratio) absolute value.bbdev: Queue Groups split into UL/DL Groups.
Queue Groups have been split into UL/DL Groups in the Turbo Software Driver. They are independent for Decode/Encode.
rte_bbdev_driver_info
reflects introduced changes.
12.4. Known Issues
Secondary process launch is not reliable.
Recent memory hotplug patches have made multiprocess startup less reliable than it was in past releases. A number of workarounds are known to work depending on the circumstances. As such it isn’t recommended to use the secondary process mechanism for critical systems. The underlying issues will be addressed in upcoming releases.
The issue is explained in more detail, including potential workarounds, in the Bugzilla entry referenced below.
Bugzilla entry: https://bugs.dpdk.org/show_bug.cgi?id=50
pdump is not compatible with old applications.
As we changed to use generic multi-process communication for pdump negotiation instead of previous dedicated unix socket way, pdump applications, including the dpdk-pdump example and any other applications using
librte_pdump
, will not work with older version DPDK primary applications.rte_abort takes a long time on FreeBSD.
DPDK processes now allocates a large area of virtual memory address space. As a result
rte_abort
on FreeBSD now dumps the contents of the whole reserved memory range, not just the used portion, to a core dump file. Writing this large core file can take a significant amount of time, causing processes to appear to hang on the system.The work around for the issue is to set the system resource limits for core dumps before running any tests, e.g.
limit coredumpsize 0
. This will effectively disable core dumps on FreeBSD. If they are not to be completely disabled, a suitable limit, e.g. 1G might be specified instead of 0. This needs to be run per-shell session, or before every test run. This change can also be made persistent by addingkern.coredump=0
to/etc/sysctl.conf
.Bugzilla entry: https://bugs.dpdk.org/show_bug.cgi?id=53
ixgbe PMD crash on hotplug detach when no VF created.
ixgbe PMD uninit path cause null pointer dereference because of port representor cleanup when number of VF is zero.
Bugzilla entry: https://bugs.dpdk.org/show_bug.cgi?id=57
Bonding PMD may fail to accept new slave ports in certain conditions.
In certain conditions when using testpmd, bonding may fail to register new slave ports.
Bugzilla entry: https://bugs.dpdk.org/show_bug.cgi?id=52.
Unexpected performance regression in Vhost library.
Patches fixing CVE-2018-1059 were expected to introduce a small performance drop. However, in some setups, bigger performance drops have been measured when running micro-benchmarks.
Bugzilla entry: https://bugs.dpdk.org/show_bug.cgi?id=48
12.6. Tested Platforms
Intel(R) platforms with Intel(R) NICs combinations
CPU
Intel(R) Atom(TM) CPU C2758 @ 2.40GHz
Intel(R) Xeon(R) CPU D-1541 @ 2.10GHz
Intel(R) Xeon(R) CPU E5-4667 v3 @ 2.00GHz
Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
Intel(R) Xeon(R) CPU E5-2699 v4 @ 2.20GHz
Intel(R) Xeon(R) CPU E5-2695 v4 @ 2.10GHz
Intel(R) Xeon(R) CPU E5-2658 v2 @ 2.40GHz
Intel(R) Xeon(R) CPU E5-2658 v3 @ 2.20GHz
Intel(R) Xeon(R) Platinum 8180 CPU @ 2.50GHz
OS:
CentOS 7.4
Fedora 25
Fedora 27
Fedora 28
FreeBSD 11.1
Red Hat Enterprise Linux Server release 7.3
SUSE Enterprise Linux 12
Wind River Linux 8
Ubuntu 14.04
Ubuntu 16.04
Ubuntu 16.10
Ubuntu 17.10
NICs:
Intel(R) 82599ES 10 Gigabit Ethernet Controller
Firmware version: 0x61bf0001
Device id (pf/vf): 8086:10fb / 8086:10ed
Driver version: 5.2.3 (ixgbe)
Intel(R) Corporation Ethernet Connection X552/X557-AT 10GBASE-T
Firmware version: 0x800003e7
Device id (pf/vf): 8086:15ad / 8086:15a8
Driver version: 4.4.6 (ixgbe)
Intel(R) Ethernet Converged Network Adapter X710-DA4 (4x10G)
Firmware version: 6.01 0x80003221
Device id (pf/vf): 8086:1572 / 8086:154c
Driver version: 2.4.6 (i40e)
Intel Corporation Ethernet Connection X722 for 10GbE SFP+ (4x10G)
Firmware version: 3.33 0x80000fd5 0.0.0
Device id (pf/vf): 8086:37d0 / 8086:37cd
Driver version: 2.4.3 (i40e)
Intel(R) Ethernet Converged Network Adapter XXV710-DA2 (2x25G)
Firmware version: 6.01 0x80003221
Device id (pf/vf): 8086:158b / 8086:154c
Driver version: 2.4.6 (i40e)
Intel(R) Ethernet Converged Network Adapter XL710-QDA2 (2X40G)
Firmware version: 6.01 0x8000321c
Device id (pf/vf): 8086:1583 / 8086:154c
Driver version: 2.4.6 (i40e)
Intel(R) Corporation I350 Gigabit Network Connection
Firmware version: 1.63, 0x80000dda
Device id (pf/vf): 8086:1521 / 8086:1520
Driver version: 5.4.0-k (igb)
Intel(R) platforms with Mellanox(R) NICs combinations
CPU:
Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz
Intel(R) Xeon(R) CPU E5-2697A v4 @ 2.60GHz
Intel(R) Xeon(R) CPU E5-2697 v3 @ 2.60GHz
Intel(R) Xeon(R) CPU E5-2680 v2 @ 2.80GHz
Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
Intel(R) Xeon(R) CPU E5-2640 @ 2.50GHz
Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.10GHz
OS:
Red Hat Enterprise Linux Server release 7.5 (Maipo)
Red Hat Enterprise Linux Server release 7.4 (Maipo)
Red Hat Enterprise Linux Server release 7.3 (Maipo)
Red Hat Enterprise Linux Server release 7.2 (Maipo)
Ubuntu 18.04
Ubuntu 17.10
Ubuntu 16.10
Ubuntu 16.04
SUSE Linux Enterprise Server 15
MLNX_OFED: 4.2-1.0.0.0
MLNX_OFED: 4.3-2.0.2.0
NICs:
Mellanox(R) ConnectX(R)-3 Pro 40G MCX354A-FCC_Ax (2x40G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1007
Firmware version: 2.42.5000
Mellanox(R) ConnectX(R)-4 10G MCX4111A-XCAT (1x10G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 10G MCX4121A-XCAT (2x10G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 25G MCX4111A-ACAT (1x25G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 25G MCX4121A-ACAT (2x25G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 40G MCX4131A-BCAT/MCX413A-BCAT (1x40G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 40G MCX415A-BCAT (1x40G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 50G MCX4131A-GCAT/MCX413A-GCAT (1x50G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 50G MCX414A-BCAT (2x50G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 50G MCX415A-GCAT/MCX416A-BCAT/MCX416A-GCAT (2x50G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 50G MCX415A-CCAT (1x100G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 100G MCX416A-CCAT (2x100G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1013
Firmware version: 12.21.1000 and above
Mellanox(R) ConnectX(R)-4 Lx 10G MCX4121A-XCAT (2x10G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1015
Firmware version: 14.21.1000 and above
Mellanox(R) ConnectX(R)-4 Lx 25G MCX4121A-ACAT (2x25G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1015
Firmware version: 14.21.1000 and above
Mellanox(R) ConnectX(R)-5 100G MCX556A-ECAT (2x100G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1017
Firmware version: 16.21.1000 and above
Mellanox(R) ConnectX-5 Ex EN 100G MCX516A-CDAT (2x100G)
Host interface: PCI Express 4.0 x16
Device ID: 15b3:1019
Firmware version: 16.21.1000 and above
ARM platforms with Mellanox(R) NICs combinations
CPU:
Qualcomm ARM 1.1 2500MHz
OS:
Red Hat Enterprise Linux Server release 7.5 (Maipo)
NICs:
Mellanox(R) ConnectX(R)-4 Lx 25G MCX4121A-ACAT (2x25G)
Host interface: PCI Express 3.0 x8
Device ID: 15b3:1015
Firmware version: 14.22.0428
Mellanox(R) ConnectX(R)-5 100G MCX556A-ECAT (2x100G)
Host interface: PCI Express 3.0 x16
Device ID: 15b3:1017
Firmware version: 16.22.0428
ARM SoC combinations from Cavium (with integrated NICs)
SoC:
Cavium CN81xx
Cavium CN83xx
OS:
Ubuntu 16.04.2 LTS with Cavium SDK-6.2.0-Patch2 release support package.
ARM SoC combinations from NXP (with integrated NICs)
SoC:
NXP/Freescale QorIQ LS1046A with ARM Cortex A72
NXP/Freescale QorIQ LS2088A with ARM Cortex A72
OS:
Ubuntu 16.04.3 LTS with NXP QorIQ LSDK 1803 support packages