6.3. Quar cluster installation

La Quar is a small town and municipality located in the comarca of Berguedà, in Catalonia.

It’s also an intel machine containing two Xilinx Alveo U200 accelerator cards.

The OmpSs-2@FPGA releases are automatically installed in the Quar cluster. They are available through a module file for each target architecture. This document describes how to load and use the modules to compile an example application. Once the modules are loaded, the workflow in the Quar cluster should be the same as in the Docker images.

6.3.1. General remarks

  • The OmpSs@FPGA toolchain is installed in a version folder under the /opt/bsc/ directory.
  • Third-party libraries required to run some programs are installed in the corresponding folder under the /opt/lib/ directory.
  • The rest of the software (Xilinx toolchain, slurm, modules, etc.) is installed under the /tools/ directory.

6.3.3. Logging into quar

Quar is accessible from HCA ssh.hca.bsc.es Alternatively, it can be accessed through the 8419 port in HCA and ssh connection will be redirected to the actual host:

ssh -p 8419 ssh.hca.bsc.es

Also, this can be automated by adding a quar host into ssh config:

Host quar
    HostName ssh.hca.bsc.es
    Port 8419

6.3.4. Module structure

The ompss-2 modules are:

  • ompss-2/x86_64/*[release version]*

This will automatically load the default Vivado version, although an arbitrary version can be loaded before ompss:

module load vivado/2023.2 ompss-2/x86_64/git

To list all available modules in the system run:

module avail

6.3.5. Build applications

To generate an application binary and bitstream, you could refer to Compile OmpSs-2@FPGA programs as the steps are general enough.

Note that the appropriate modules need to be loaded. See Module structure.

6.3.6. Running applications

Get access to an installed fpga

Quar cluster uses SLURM in order to manage access to computation resources. Therefore, to be able to use the resources of an FPGA, an allocation in one of the partitions has to be made.

There is 1 partition in the cluster:

  • fpga: two Alveo U200 boards

The easiest way to allocate an FPGA is to run bash through srun with the --gres option:

srun --gres=fpga:BOARD:N --pty bash

Where BOARD is the FPGA to allocate, in this case alveo_u200, and N the number of FPGAs to allocate, either 1 or 2.

For instance, the command:

srun --gres=fpga:alveo_u200:2 --pty bash

Will allocate both FPGAs and run an interactive bash with the required tools and file permissions already set by slurm. To get information about the active slurm jobs, run:

squeue

The output should look similar to this:

JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
 1312 fpga          bash afilguer  R      17:14      1 quar

Loading bistreams

The FPGA bitstream needs to be loaded before the application can run. The load_bitstream utility is provided in order to simplify the FPGA configuration.

load_bitstream bitstream.bit [index]

The utility receives a second optional parameter to indicate which of the allocated FPGAs to program, the default behavior is to program all the allocated FPGAs with the bitstream.

To know which FPGAs indices have been allocated, run load_bitstream with the help (-h) option. The output should be similar to this:

Usage load_bitstream bitstream.bit [index]
Available devices:
index:  jtag           pcie          usb
0:      21290594G00LA  0000:b3:00.0  001:002
1:      21290594G00EA  0000:65:00.0  001:003

Set up qdma queues

Note

This step is performed by load_bitstream script, which creates a single bidirectional memory mapped queue. This is only needed if other configuration is needed.

For DMA transfers to be performed between system main memory and the FPGA memory, qdma queues has to be set up by the user prior to any execution.

In this case dmactl tool is used. For instance: In order to create and start a memory mapped qdma queue with index 1 run:

dmactl qdmab3000 q add idx 1 mode mm dir bi
dmactl qdmab3000 q start idx 1 mode mm dir bi

OmpSs runtime system expects an mm queue at index 1, which can be created with the commands listed above.

In the same fashion, these queues can also be removed:

dmactl qdmab3000 q stop idx 1 mode mm dir bi
dmactl qdmab3000 q del idx 1 mode mm dir bi

For more information, see

dmactl --help

Get current bitstream info

In order to get information about the bitstream currently loaded into the FPGA, the tool read_bitinfo is installed in the system.

read_bitinfo

Note that an active slurm reservation is needed in order to query the FPGA.

This call should return something similar to the sample output for a cholesky decomposition application:

Bitinfo of FPGA 0000:b3:00.0:
Bitinfo version:    13
Bitstream user-id:  0x603186FD
AIT version:        7.7.1
Wrapper version     13
Number of acc:      9
Board base frequency (MHz)  156.250000
Interleaving stride 32768

Features:
[ ] Instrumentation
[ ] Hardware counter
[x] Performance interconnect
[ ] Simplified interconnection
[ ] POM AXI-Lite
[x] POM task creation
[x] POM dependencies
[ ] POM lock
[x] POM spawn queues
[ ] Power monitor (CMS)
[ ] Thermal monitor (sysmon)
[ ] OMPIF

Managed rstn addr 0xA000
Cmd In addr 0x6000 len 64
Cmd Out addr 0x8000 len 64
Spawn In addr 0x2000 len 1024
Spawn Out addr 0x4000 len 1024
Hardware counter not enabled
POM AXI-Lite not enabled
Power monitor (CMS) not enabled
Thermal monitor (sysmon) not enabled

xtasks accelerator config:
type        count   freq(KHz)   description
5862896218  1       300000      cholesky_blocked
5459683074  1       300000      omp_trsm
5459653839  1       300000      omp_syrk
5459186490  6       300000      omp_gemm

ait command line:
ait --name=cholesky --board=alveo_u200 -c=300 --max_deps_per_task=3 --max_args_per_task=3 --max_copies_per_task=3 --picos_tm_size=256 --picos_dm_size=645 --picos_vm_size=775 --memory_interleaving_stride=32K --simplify_interconnection --interconnect_priorities --interconnect_opt=performance --interconnect_regslice=all --floorplanning_constr=acc --slr_slices=static --placement_file=u200_placement_6x256.json --disable_creator_ports --wrapper_version 13

Hardware runtime VLNV:
bsc:ompss:picos_ompss_manager:7.3

Remote debugging

Although it is possible to interact with Vivado’s Hardware Manager through ssh-based X forwarding, Vivado’s GUI might not be very responsive over remote connections. To avoid this limitation, one might connect a local Hardware Manager instance to targets hosted on Quar, completely avoiding X forwarding, as follows.

  1. On Quar, when allocating an FPGA with slurm, a Vivado HW server is automatically launched for each FPGA:
    • FPGA 0 uses port 3120
    • FPGA 1 uses port 3121
  2. On the local machine, assuming that Quar’s HW Server runs on port 3120, let all connections to port 3120 be forwarded to quar by doing ssh -L 3120:quar:3120 [USER]@ssh.hca.bsc.es -p 8410 .
  3. Finally, from the local machine, connect to Quar’s hardware server:
    • Open Vivado’s Hardware Manager.
    • Launch the “Open target” wizard.
    • Establish a connection to the local HW server, which will be just a bridge to the remote instance.

Enabling OpenCL / XRT mode

FPGA in quar can be used in OpenCL / XRT mode. Currently, XRT 2022.2 is installed. To enable XRT the shell has to be configured into the FPGA and the PCIe devices re-enumerated after configuration has finished.

This is done by running

load_xrt_shell

Note that this has to be done while a slurm job is allocated. After this process has completed, output from lspci -vd 10ee: should look similar to this:

b3:00.0 Processing accelerators: Xilinx Corporation Device 5000
        Subsystem: Xilinx Corporation Device 000e
        Flags: bus master, fast devsel, latency 0, NUMA node 0
        Memory at 383ff0000000 (64-bit, prefetchable) [size=32M]
        Memory at 383ff4000000 (64-bit, prefetchable) [size=256K]
        Capabilities: <access denied>
        Kernel driver in use: xclmgmt
        Kernel modules: xclmgmt

b3:00.1 Processing accelerators: Xilinx Corporation Device 5001
        Subsystem: Xilinx Corporation Device 000e
        Flags: bus master, fast devsel, latency 0, IRQ 105, NUMA node 0
        Memory at 383ff2000000 (64-bit, prefetchable) [size=32M]
        Memory at 383ff4040000 (64-bit, prefetchable) [size=256K]
        Memory at 383fe0000000 (64-bit, prefetchable) [size=256M]
        Capabilities: <access denied>
        Kernel driver in use: xocl
        Kernel modules: xocl

Also XRT devices should show up as ready when running xbutil examine. Note that the xrt/2022.2 has to be loaded.

module load xrt/2022.2
xbutil examine

And it should show this output:

System Configuration
  OS Name              : Linux
  Release              : 5.4.0-97-generic
  Version              : #110-Ubuntu SMP Thu Jan 13 18:22:13 UTC 2022
  Machine              : x86_64
  CPU Cores            : 16
  Memory               : 63812 MB
  Distribution         : Ubuntu 18.04.2 LTS
  GLIBC                : 2.31
  Model                : PowerEdge T640

XRT
  Version              : 2.14.354
  Branch               : 2022.2
  Hash                 : 43926231f7183688add2dccfd391b36a1f000bea
  Hash Date            : 2022-10-08 09:49:58
  XOCL                 : 2.14.354, 43926231f7183688add2dccfd391b36a1f000bea
  XCLMGMT              : 2.14.354, 43926231f7183688add2dccfd391b36a1f000bea

Devices present
BDF             :  Shell                            Platform UUID                         Device ID       Device Ready*
-------------------------------------------------------------------------------------------------------------------------
[0000:b3:00.1]  :  xilinx_u200_gen3x16_xdma_base_2  0B095B81-FA2B-E6BD-4524-72B1C1474F18  user(inst=128)  Yes


* Devices that are not ready will have reduced functionality when using XRT tools