Batch Sealing with SupraSeal

This page explains how to setup supraseal batch sealer in Curio

Disclaimer: SupraSeal batch sealing is currently in BETA. Use with caution and expect potential issues or changes in future versions. Currently some additional manual system configuration is required. Batch Sealing only supports "CC" sectors as of now. Please make sure that "SnapDeals" are enabled in the cluster if you wish to onboard data with SupraSeal enabled. If SnapDeals are not enabled, deals will be routed to SupraSeal pipeline which will discard the actual data and seal empty sectors.

SupraSeal is an optimized batch sealing implementation for Filecoin that allows sealing multiple sectors in parallel. It can significantly improve sealing throughput compared to sealing sectors individually.

Key Features

Seals multiple sectors (up to 128) in a single batch
- Up to 16x better core utilisation efficiency
Optimized to utilize CPU and GPU resources efficiently
Uses raw NVMe devices for layer storage instead of RAM

Requirements

CPU with at least 4 cores per CCX (AMD) or equivalent
NVMe drives with high IOPS (10-20M total IOPS recommended)
GPU for PC2 phase (NVIDIA RTX 3090 or better recommended)
1GB hugepages configured (minimum 36 pages)
Ubuntu 22.04 or compatible Linux distribution (gcc-11 required, doesn't need to be system-wide)
At least 256GB RAM, ALL MEMORY CHANNELS POPULATED
- Without all memory channels populated sealing performance will suffer drastically
NUMA-Per-Socket (NPS) set to 1

Storage Recommendations

You need 2 sets of NVMe drives:

Drives for layers:
- Total 10-20M IOPS
- Capacity for 11 x 32G x batchSize x pipelines
- Raw unformatted block devices (SPDK will take them over)
- Each drive should be able to sustain ~2GiB/s of writes
  - This requirement isn't understood well yet, it's possible that lower write rates are fine. More testing is needed.
Drives for P2 output:
- With a filesystem
- Fast with sufficient capacity (~70G x batchSize x pipelines)
- Can be remote storage if fast enough (~500MiB/s/GPU)

The following table shows the number of NVMe drives required for different batch sizes. The drive count column indicates N + M where N is the number of drives for layer data (SPDK) and M is the number of drives for P2 output (filesystem). The iops/drive column shows the minimum iops per drive required for the batch size. Batch size indicated with 2x means dual-pipeline drive setup. IOPS requirements are calculated simply by dividing total target 10M IOPS by the number of drives. In reality, depending on CPU core speed this may be too low or higher than neccesary. When ordering a system with barely enough IOPS plan to have free drive slots in case you need to add more drives later.

Batch Size

3.84TB

7.68TB

12.8TB

15.36TB

30.72TB

4 + 1

2 + 1

1 + 1

^ iops/drive

2500K

5000K

10000K

64 (2x 32)

7 + 2

4 + 1

2 + 1

1 + 1

^ iops/drive

1429K

2500K

5000K

10000K

128 (2x 64)

13 + 3

7 + 2

4 + 1

2 + 1

^ iops/drive

770K

1429K

2500K

5000K

2x 128

26 + 6

13 + 3

8 + 2

7 + 2

4 + 1

^ iops/drive

385K

770K

1250K

1429K

2500K

Hardware Recommendations

Currently, the community is trying to determine the best hardware configurations for batch sealing. Some general observations are:

Single socket systems will be easier to use at full capacity
You want a lot of NVMe slots, on PCIe Gen4 platforms with large batch sizes you may use 20-24 3.84TB NVMe drives
In general you'll want to make sure all memory channels are populated
You need 4~8 physical cores (not threads) for batch-wide compute, then on each CCX you'll lose 1 core for a "coordinator"
- Each thread computes 2 sectors
- On zen2 and earlier hashers compute only one sector per thread
- Large (many-core) CCX-es are typically better

Please consider contributing to the SupraSeal hardware examples.

Setup

Check NUMA setup:

numactl --hardware

You should expect to see available: 1 nodes (0). If you see more than one node you need to go into your UEFI and set NUMA Per Socket (or a similar setting) to 1.

Configure hugepages:

This can be done by adding the following to /etc/default/grub. You need 36 1G hugepages for the batch sealer.

GRUB_CMDLINE_LINUX_DEFAULT="hugepages=36 default_hugepagesz=1G hugepagesz=1G"

Then run sudo update-grub and reboot the machine.

Or at runtime:

sudo sysctl -w vm.nr_hugepages=36

Then check /proc/meminfo to verify the hugepages are available:

cat /proc/meminfo | grep Huge

Expect output like:

AnonHugePages:         0 kB
ShmemHugePages:        0 kB
FileHugePages:         0 kB
HugePages_Total:      36
HugePages_Free:       36
HugePages_Rsvd:        0
HugePages_Surp:        0
Hugepagesize:    1048576 kB

Check that HugePages_Free is equal to 36, the kernel can sometimes use some of the hugepages for other purposes.

Dependencies

CUDA 12.x is required, 11.x won't work. The build process depends on GCC 11.x system-wide or gcc-11/g++-11 installed locally.

On Arch install https://aur.archlinux.org/packages/gcc11
Ubuntu 22.04 has GCC 11.x by default
On newer Ubuntu install gcc-11 and g++-11 packages
```
sudo apt install gcc-11 g++-11
```
In addtion to general build dependencies (listed on the installation page), you need libgmp-dev and libconfig++-dev
```
sudo apt install libgmp-dev libconfig++-dev
```

Building

Build and install the batch-capable Curio binary:

make batch
make sptool

make install

For calibnet

make batch-calibnet
make batch-sptool

make install

The build should be run on the target machine. Binaries won't be portable between CPU generations due to different AVX512 support.

Setup NVMe devices for SPDK:

This is only needed while batch sealing is in beta, future versions of Curio will handle this automatically.

cd extern/supraseal/deps/spdk-v24.05/
env NRHUGE=36 ./scripts/setup.sh

Benchmark NVME IOPS

Please make sure to benchmark the raw NVME IOPS before proceeding with further configuration to verify that IOPS requirements are fulfilled.

cd extern/supraseal/deps/spdk-v24.05/

# repeat -b with all devices you plan to use with supraseal
# NOTE: You want to test with ALL devices so that you can see if there are any bottlenecks in the system
./build/examples/perf -b 0000:85:00.0 -b 0000:86:00.0...  -q 64 -o 4096 -w randread -t 10

The output should look like below

========================================================
                                                                           Latency(us)
Device Information                     :       IOPS      MiB/s    Average        min        max
PCIE (0000:04:00.0) NSID 1 from core  0:  889422.78    3474.31      71.93      10.05    1040.94
PCIE (0000:41:00.0) NSID 1 from core  0:  890028.08    3476.67      71.88      10.69    1063.32
PCIE (0000:42:00.0) NSID 1 from core  0:  890035.08    3476.70      71.88      10.66    1001.86
PCIE (0000:86:00.0) NSID 1 from core  0:  889259.28    3473.67      71.95      10.62    1003.83
PCIE (0000:87:00.0) NSID 1 from core  0:  889179.58    3473.36      71.95      10.55     993.32
PCIE (0000:88:00.0) NSID 1 from core  0:  889272.18    3473.72      71.94      10.38     974.63
PCIE (0000:c1:00.0) NSID 1 from core  0:  889815.08    3475.84      71.90      10.97    1044.70
PCIE (0000:c2:00.0) NSID 1 from core  0:  889691.08    3475.36      71.91      11.04    1036.57
PCIE (0000:c3:00.0) NSID 1 from core  0:  890082.78    3476.89      71.88      10.44    1023.32
========================================================
Total                                  : 8006785.90   31276.51      71.91      10.05    1063.32

With ideally >10M IOPS total for all devices.

PC2 output storage

Attach scratch space storage for PC2 output (batch sealer needs ~70GB per sector in batch - 32GiB for the sealed sector, and 36GiB for the cache directory with TreeC/TreeR and aux files)

Usage

Start the Curio node with the batch sealer layer

curio run --layers batch-machine1

Add a batch of CC sectors:

curio seal start --now --cc --count 32 --actor f01234 --duration-days 365

Monitor progress - you should see a "Batch..." task running in the Curio GUI
PC1 will take 3.5-5 hours, followed by PC2 on GPU
After batch completion, the storage will be released for the next batch

Configuration

Run curio calc batch-cpu on the target machine to determine supported batch sizes for your CPU

Example batch-cpu output

# EPYC 7313 16-Core Processor

root@udon:~# ./curio calc batch-cpu
Basic CPU Information

Processor count: 1
Core count: 16
Thread count: 32
Threads per core: 2
Cores per L3 cache (CCX): 4
L3 cache count (CCX count): 4
Hasher Threads per CCX: 6
Sectors per CCX: 12
---------
Batch Size: 16 sectors

Required Threads: 8
Required CCX: 2
Required Cores: 6 hasher (+4 minimum for non-hashers)
Enough cores available for hashers ✔
Non-hasher cores: 10
Enough cores for coordination ✔

pc1 writer: 1
pc1 reader: 2
pc1 orchestrator: 3

pc2 reader: 4
pc2 hasher: 5
pc2 hasher_cpu: 6
pc2 writer: 7
pc2 writer_cores: 3

c1 reader: 7

Unoccupied Cores: 0

{
  sectors = 16;
  coordinators = (
    { core = 10;
      hashers = 2; },
    { core = 12;
      hashers = 6; }
  )
}
---------
Batch Size: 32 sectors

Required Threads: 16
Required CCX: 3
Required Cores: 11 hasher (+4 minimum for non-hashers)
Enough cores available for hashers ✔
Non-hasher cores: 5
Enough cores for coordination ✔
! P2 hasher will share a core with P1 writer, performance may be impacted
! P2 hasher_cpu will share a core with P2 reader, performance may be impacted

pc1 writer: 1
pc1 reader: 2
pc1 orchestrator: 3

pc2 reader: 0
pc2 hasher: 1
pc2 hasher_cpu: 0
pc2 writer: 4
pc2 writer_cores: 1

c1 reader: 0

Unoccupied Cores: 0

{
  sectors = 32;
  coordinators = (
    { core = 5;
      hashers = 4; },
    { core = 8;
      hashers = 6; },
    { core = 12;
      hashers = 6; }
  )
}
---------
Batch Size: 64 sectors

Required Threads: 32
Required CCX: 6
Required Cores: 22 hasher (+4 minimum for non-hashers)
Not enough cores available for hashers ✘
Batch Size: 128 sectors

Required Threads: 64
Required CCX: 11
Required Cores: 43 hasher (+4 minimum for non-hashers)
Not enough cores available for hashers ✘

Create a new layer configuration for the batch sealer, e.g. batch-machine1:

[Subsystems]
EnableBatchSeal = true

[Seal]
LayerNVMEDevices = [
  "0000:88:00.0",
  "0000:86:00.0", 
  # Add PCIe addresses for all NVMe devices to use
]

# Set to your desired batch size (what the batch-cpu command says your CPU supports AND what you have nvme space for)
BatchSealBatchSize = 32

# pipelines can be either 1 or 2; 2 pipelines double storage requirements but in correctly balanced systems makes
# layer hashing run 100% of the time, nearly doubling throughput
BatchSealPipelines = 2

# Set to true for Zen2 or older CPUs for compatibility
SingleHasherPerThread = false

Optimization

Balance batch size, CPU cores, and NVMe drives to keep PC1 running constantly
Ensure sufficient GPU capacity to complete PC2 before next PC1 batch finishes
Monitor CPU, GPU and NVMe utilization to identify bottlenecks
Monitor hasher core utilisation

Troubleshooting

Node doesn't start / isn't visible in the UI

Ensure hugepages are configured correctly
Check NVMe device IOPS and capacity
- If spdk setup fails, try to wipefs -a the NVMe devices (this will wipe partitions from the devices, be careful!)

Performance issues

You can monitor performance by looking at "hasher" core utilisation in e.g. htop.

To identify hasher cores, call curio calc supraseal-config --batch-size 128 (with the correct batch size), and look for coordinators

topology:
...
{
  pc1: {
    writer       = 1;
...
    hashers_per_core = 2;

    sector_configs: (
      {
        sectors = 128;
        coordinators = (
          { core = 59;
            hashers = 8; },
          { core = 64;
            hashers = 14; },
          { core = 72;
            hashers = 14; },
          { core = 80;
            hashers = 14; },
          { core = 88;
            hashers = 14; }
        )
      }

    )
  },

  pc2: {
...
}

In this example, cores 59, 64, 72, 80, and 88 are "coordinators", with two hashers per core, meaning that

In first group core 59 is a coordinator, cores 60-63 are hashers (4 hasher cores / 8 hasher threads)
In second group core 64 is a coordinator, cores 65-71 are hashers (7 hasher cores / 14 hasher threads)
And so on

Coordinator cores will usually sit at 100% utilisation, hasher threads SHOULD sit at 100% utilisation, anything less indicates a bottleneck in the system, like not enough NVMe IOPS, not enough Memory bandwidth, or incorrect NUMA setup.

To troubleshoot:

Read the requirements at the top of this page very carefully
Benchmark NVME IOPS
Validate GPU setup if PC2 is slow
Review logs for any errors during batch processing

Slower than expeted NVMe speed

If the NVME Benchmark shows lower than expected IOPS, you can try formatting the NVMe devices with SPDK:

cd extern/supraseal/deps/spdk-v24.05/
./build/examples/nvme_manage

Go through the menus like this

NVMe Management Options
	[1: list controllers]
	[2: create namespace]
	[3: delete namespace]
	[4: attach namespace to controller]
	[5: detach namespace from controller]
	[6: format namespace or controller]
	[7: firmware update]
	[8: opal]
	[9: quit]
6

0000:04:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829861           6 
0000:41:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829862           6 
0000:42:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829860           6 
0000:86:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829794           6 
0000:87:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829798           6 
0000:88:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829837           6 
0000:c1:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829795           6 
0000:c2:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829836           6 
0000:c3:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829797           6 
0000:c4:00.00 SAMSUNG MZQL23T8HCLS-00A07               S64HNG0W829850           6 
Please Input PCI Address(domain:bus:dev.func):
0000:c4:00.00
Please Input Namespace ID (1 - 32):
1                                                          ## Select 1

Please Input Secure Erase Setting:
	0: No secure erase operation requested
	1: User data erase
	2: Cryptographic erase
0

Supported LBA formats:
 0: 512 data bytes
 1: 4096 data bytes
Please input LBA format index (0 - 1):
1                                                          ## Select 4096 data bytes

Warning: use this utility at your own risk.
This command will format your namespace and all data will be lost.
This command may take several minutes to complete,
so do not interrupt the utility until it completes.
Press 'Y' to continue with the format operation.
y

Then you might see a difference in performance like this:

                                                                           Latency(us)
Device Information                     :       IOPS      MiB/s    Average        min        max
PCIE (0000:c1:00.0) NSID 1 from core  0:  721383.71    2817.91      88.68      11.20     591.51  ## before
PCIE (0000:86:00.0) NSID 1 from core  0: 1205271.62    4708.09      53.07      11.87     446.84  ## after

PreviousSnap Deals NextScaling Curio cluster

Last updated 2 months ago