The IBM POWER7 big endian systems, cfarm110 and cfarm111, will be decommissioned to simplify the OSUOSL consolidation and transition.  A big-endian partition on a POWER10 system remains available as cfarm121.

The systems will be turned off on Friday, February 27.  Please ensure that you copy any data that you wish to preserve from those systems before that date.  None of the data on CFarm systems is backed up and none of the data will be recoverable after that date.

The CFarm project wishes to thank IBM for providing the systems and OSUOSL for hosting the systems.  When initially contributed, they were some of the most powerful systems in the CFarm.  They have been extremely valuable to FOSS community developers for over a decade.

The Compile Farm project is happy to announce the immediate availability of cfarm107 and cfarm108, two identical NVIDIA DGX Spark aarch64 machines based on the Grace Blackwell GB10 graphics processing chip running DGX OS 7 (Ubuntu 24.04, R580 drivers, CUDA 13.0) as of writing.

Please note the non-standard SSH ports, 2107 and 2108, respectively. Standard ports and inbound IPv6 access may be added in the future. (Outbound IPv6 is available).

This system-on-chip development platform features a 20-core heterogeneous big.LITTLE ARM Cortex-X925/A725 package (10 cores each), 128GB of LPDDR5X-8533 unified system memory claiming 273 GB/s, and a 4TB PCIe Gen.5 NVMe.

Users should be able to follow the DGX Spark Playbooks with ease, however 4TB will disappear quickly if we’re not careful to delete old data and avoid hoarding large files. Some features may not be available today.

Please see the cfarm-users mailing list for additional details and discussion about these machines as we work with users to configure them for different workflows.

Thank you to NVIDIA and David Edelsohn for sponsoring these machines, and to Adélie Linux for hosting and maintaining them!

cfarm202, one of our Debian SPARC systems, had been offline for some time due to an hardware issue.

The machine has now been repaired with a new SPARC CPU module donated by Jeffrey Walton (thanks!). Jeffrey also donated a second CPU module to speed up repairs in case of future similar failure.

In addition, partial disk backups from 2023 have been restored on the machine. Not all home directories could be restored.

As usual, disk space is a shared resource, so please remember to clean up any old files in your home directory.

Due to hardware issues, cfarm203 had to be decommisionned. This was the only big-endian POWER machine running Debian in the farm.

To replace it, cfarm121 has been installed with a big-endian Debian OS. While the old machine was a POWER8 (IBM 8284-22A), the new machine is a POWER10 (more precisely, a KVM-based virtual machine on a IBM 9105-42A).

Home directories from cfarm203 have been restored to cfarm121, but unfortunately only a backup from 2023 could be used. This is a reminder to do your own backups if you have important data on farm machines; additionally, unused data should be removed to keep disk usage manageable (this is a shared resource).

Many thanks to John Paul Adrian Glaubitz for driving this replacement and to OSUOSL for hosting the new machine.

We are happy to announce the addition of cfarm95, a BPI-F3 machine from BananaPi. This RISC-V machine is based on a SpacemiT K1 SoC with 8 SpacemiT X60 cores. The most interesting feature of this hardware is its implementation of the RVV 1.0 vector extension; it also conforms to the RVA22 standard. More hardware information.

The machine runs a standard Debian userspace; however, the kernel is a vendor kernel derived from Linux 6.6. Once upstream kernel support for this board becomes available, we hope to switch to the upstream kernel to improve stability. In the meantime, we have documented how to run upstream Debian on this hardware.

This machine should help free software developers add support for RISC-V vector extensions in their projects. This is especially interesting for toolchains and performance-sensitive code such as cryptographic libraries.

Thanks to RISC-V International for providing the hardware, and to tetaneutral.net for hosting it!

We are happy to announce the immediate availability of cfarm423, cfarm424, cfarm425, and cfarm426, all of which are aarch64 systems. They are hosted in Tokyo, Japan.

cfarm423 is a Radxa ROCK 5B, running Debian 13 trixie, with upstream u-boot and a custom (64K PAGESIZE) linux build that follows Debian 13's release.

The other 3 systems are virtual machines based on an Ampere Altra Max M128-30 system. Each VM has 64 cores and 64GB of memory and runs a different OS:

cfarm424: Debian 13 trixie
cfarm425: Debian testing-forever (currently this is Debian 13, but it will move to Debian 14 and so on)
cfarm426: Rocky 9

Storage for /home is plenty, all are on NVMe with ZFS's block level compression enabled: 4TB for cfarm423, 3TB each for the 3 VMs cfarm424~426.

This addition to the farm should provide good build times for highly parallel workloads on aarch64, and will provide low latency for users in Asia. Enjoy!

We are very happy to announce the immediate availability of two new RISC-V hosts: cfarm93 running upstream Debian, and cfarm94 running upstream Alpine Linux. Both hosts are StarFive VisionFive 2 boards, which is currently the RISC-V hardware with the best balance of upstream software support and performance. We use a fully upstream version of u-boot as well as a near-upstream Linux kernel: the only addition to the upstream kernel are PCIe patches that are pending upstream integration, so that we can use a NVMe drive on the boards.

The hardware was sponsored by RISC-V International, while NVMe drives and hosting are provided by tetaneutral.net. Many thanks to both organizations for their support.

The farm already provides experimental RISC-V hosts since July 2022: cfarm91 (VisionFive 1) and cfarm92 (HiFive Unmatched). However, they clearly provide much less performance than the new boards, and their hardware is no longer produced. We will keep these old hosts online on a best-effort basis while the hardware is working, but without much expectations (e.g. without OS upgrades).

We are pleased to announce the availability of two new machines, cfarm215 and cfarm216. They run the latest Solaris 11.4 SRU (Support Repository Update). cfarm215 is a kernel zone hosted on a Dell R740 system with an Intel Xeon CPU, while cfarm216 is a LDOM hosted on an Oracle SPARC T8-1 with a SPARC M8 CPU.

Basic development packages are installed from the Solaris support repository. More packages can be installed on request.

Thanks to the Center for Biotechnology (CeBiTec) at Bielefeld University in Germany for setting up and hosting these machines. The R740 is a donation from de.NBI Cloud Bielefeld, while the T8-1 is a permanent loan from Oracle Corporation.

We have added three new x86_64 virtual machines based on a dual AMD EPYC 7773X system with plenty of CPU cores. Each VM has 128 logical cores + 64 GB of memory and runs a different OS:

cfarm420 runs Arch Linux

cfarm421 runs Debian 13 trixie

cfarm422 runs Debian 12 bookworm

This addition to the farm should provide good build times for highly parallel workloads, and will provide low latency for users in Asia.

The virtual machines are located in Tokyo, Japan. Thanks to Jing Luo for providing the hardware, hosting and support!

TL;DR: on cfarm23, your old home data is in /oldhome, please copy anything useful before it becomes unavailable.

Our smaller hosts usually can't have a local disk, so in that case we provide home directories over NFS from a bigger server.

This has been the case for cfarm23, one of our EdgeRouter host. Unfortunately, due to hardware constraints, we have to change the disk backing its home directory and we cannot migrate the data.

As a result, cfarm23 now has a brand new empty home directory, mounted over NFS from a new server. We still provide the old home mounted on /oldhome, but cannot make any promise about how long this old data will stay available. Please copy any useful data from the old home before it becomes permanently unavailable.

Additionally, cfarm91, our RISC-V VisionFive1 host, now also uses a NFS home instead of the local SD card. This should speed up builds significantly, although the CPU cores remain quite slow. In that case, we were able to copy the data over, which means that no action is required.