Merge pull request #1230 from Zarquan/20231029-zrq-rsap-request

Updating RSAP requests

notes/zrq/20231029-01-rsap-request.txt

@@ -0,0 +1,245 @@
+#
+# <meta:header>
+#   <meta:licence>
+#     Copyright (c) 2023, ROE (http://www.roe.ac.uk/)
+#
+#     This information is free software: you can redistribute it and/or modify
+#     it under the terms of the GNU General Public License as published by
+#     the Free Software Foundation, either version 3 of the License, or
+#     (at your option) any later version.
+#
+#     This information is distributed in the hope that it will be useful,
+#     but WITHOUT ANY WARRANTY; without even the implied warranty of
+#     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+#     GNU General Public License for more details.
+#
+#     You should have received a copy of the GNU General Public License
+#     along with this program.  If not, see <http://www.gnu.org/licenses/>.
+#   </meta:licence>
+# </meta:header>
+#
+#zrq-notes-time
+#zrq-notes-indent
+#zrq-notes-crypto
+#zrq-notes-ansible
+#zrq-notes-osformat
+#zrq-notes-zeppelin
+#
+# AIMetrics: []
+#
+
+    Target:
+
+        2023/24/25 RSAP request
+
+    Result:
+
+        Work in progress ...
+
+# -----------------------------------------------------
+
+    Cambridge
+
+        Live DR4 service
+              700 cores
+            1 TiB RAM
+        Full DR4 dataset, 500 TiB
+           50 TiB Object Swift
+          500 TiB CephFS Manila
+            5 TiB Block  Cinder
+           16 TiB Direct attached SSD Ephemeral
+
+        Dev DR4 service
+              700 cores
+            1 TiB RAM
+        Full DR4 dataset, 500 TiB (needed to re-generate table indexing)
+           50 TiB Object Swift
+          500 TiB CephFS Manila
+            5 TiB Block  Cinder
+           16 TiB Direct attached SSD Ephemeral
+
+        Total for Cambridge
+             1400 cores
+            2 TiB RAM
+          100 TiB Object Swift
+         1000 TiB CephFS Manila
+           10 TiB Block  Cinder
+           32 TiB Direct attached SSD Ephemeral
+
+    Somerville
+
+        Live/dev DR4 service
+              700 cores
+            1 TiB RAM
+        Full DR4 dataset, 500 TiB
+           50 TiB Object Swift
+          500 TiB CephFS Manila
+            5 TiB Block  Cinder
+           16 TiB Direct attached SSD Ephemeral
+
+    STFC cloud RAL
+
+        Live/dev DR4 service
+              700 cores
+            1 TiB RAM
+        Full DR4 dataset, 500 TiB
+           50 TiB Object Swift
+          500 TiB CephFS Manila
+            5 TiB Block  Cinder
+           16 TiB Direct attached SSD Ephemeral
+
+# -----------------------------------------------------
+
+    RSAP total
+
+             2800 cores
+            4 TiB RAM
+          200 TiB Object Swift
+         2000 TiB CephFS Manila
+           20 TiB Block  Cinder
+           64 TiB Direct attached SSD Ephemeral
+
+# -----------------------------------------------------
+
+    Peak estimates:
+
+        This resource request is based on estimates for the peak load
+        expected when Gaia DR4 is released at the end of 2025.
+
+        Our current Gaia DR3 dataset is around 8TiB.
+        The Gaia DR4 dataset is expected be be in the order of 500TiB.
+
+        The current timeline for Gaia is to release Gaia DR4 in Q4 2025,
+        although we may start to receive pre-release copies of the data
+        for internal development and testing during 2024/25.
+
+        When DR4 is released in Q4 2025 we are expecting to have to curate multiple
+        copies of the 500TiB dataset, and handle a peak of interest from users in
+        response to the publicity surrounding the event.
+
+        This means that our resource requirements for this 2024/2025 RSAP round
+        includes an estimate of the resources that will need to be deployed by
+        the end of 2024, ready to handle the peak load leading up to the DR4
+        release at the end of 2025.
+
+        Resources requested in the 2025/2026 RSAP round will not be in deployed,
+        tested and ready for use in time.
+
+        Once our Gaia DR4 platform has been deployed and the expected peak load
+        has passed; we may be able to release some of the requested resources.
+
+    Multiple sites:
+
+        We are planning for our main live deployment to be at Cambridge, along
+        with a second system on the same platform for development and data curation.
+
+        The secondary sites at RAL and Somerville will provide scale-out capacity to
+        handle peak load expected during the DR4 release.
+
+        IF secondary sites at RAL and Somerville come online with no problems,
+        then we may be able to release some the second system at Cambridge.
+
+        Our development plan for 2024 includes migrating from the current monolithic
+        deployment to a more flexible system consisting of an initial website
+        that handles user accounts and login, backed by a number of notebook services.
+
+        This will enable us to scale out to use multiple physical sites for the
+        notebook services while presenting the user with what appears to be a single
+        integrated platform.
+
+    Multiple data copies:
+
+        Our development plan for 2024 includes working with mock copies of the DR4
+        dataset to test our data injest, indexing and partitioning processes,
+        and testing new deployment methods that can scale out to cope with this
+        amount of data.
+
+        During 2025 we expect to be injesting pre-release versions of the DR4 data
+        set and using them to run a beta-test version of the system for a restricted
+        set of users, while at the same time running one or more development sites,
+        and keeping the existing live DR3 site running.
+
+    Booking system:
+
+        Our development plan for 2024 includes a booking system that enables users
+        to book sessions ahead of time, smoothing out the peaks in demand and enabling
+        us to predict the resources needed in advance.
+
+        With this in place, we may be able to release resources at some of the physical
+        sites in response to changes in load.
+
+        It is hoped that the booking system can be integrated with Blazar, the Openstack
+        resource reservation system, if it becomes available on the Openstack platforms.
+
+    Resource pinning:
+
+        Our current allocation at Cambridge is pinned to specific hardware.
+        This is to needed guarantee that we can release and create sets of resources without
+        running into problems.
+
+        In the past we found that if the Openstack system is configured to make maximal
+        use all the available resources, leaving few idle resources available for new
+        allocations, then releasing and creating a block of 10+ VMs may fail because some
+        of the resources get assigned to other processes during the gap between our
+        release and create steps.
+
+        Whether this will be needed at RAL or Somerville will depend on the configuration
+        of their Openstack systems.
+
+        This issue may be solved by the deployment of Blazar, the Openstack resource
+        reservation system.
+
+    High memory hypervisors:
+
+        A number of the hypervisors at Cambridge are high memory nodes with a higher
+        memory:vcpu ratio.
+
+        These host the high memory VMs required by the Zeppelin head nodes to support
+        the in-memory processing used by libraries such as HDBSCAN.
+        (*) cite Dennis's paper ?
+
+    Project specific flavors:
+
+        We have a set of project specific flavors at Cambridge, designed to optimise
+        the packing of VMs into the pinned hypervisors.
+
+        This includes a set of 'himem' flavors designed to make use of the high-memory
+        hypervisors.
+
+        We will also need to create a set of 'high storage' flavors to make use of the
+        direct attached SSD storage when it becomes available.
+
+    Direct attached storage:
+
+        This request is based on the assumption that direct attached SSD storage wil
+        be able to solve the problems with high IO wait that we have seen with our
+        current deployment.
+        We planned to use the SSD resources requested in our 2023 allocation to test
+        this theory, but the resources are not available yet.
+
+        We hope to be able to resolve this question during work planned for 2024.
+
+    CephFS vs object storage:
+
+        The current system uses the CephFS storage system at Cambridge, which is is based
+        on spinning disc hard drives.
+
+        In theory it should be possible to serve the Gaia dataset via the S3 interface
+        of Ceph object storage using the Openstack Swift interface.
+        However, we have encountered some issues with the Java S3 client libraries
+        when using this to serve the Gaia DR3 dataset.
+        Which is why this request specifies the majority of the storage, 500TiB, on
+        CephFS via Manila and only 50 TiB of object via Swift at each site.
+
+           50 TiB Object Swift
+          500 TiB CephFS Manila
+
+        If these issues can be solved, and if the resulting S3 interface is fast enough
+        to provide the bandwidth needed, then we may be able to swap the bulk of the
+        storage, 500TiB, from CephFS via Manila to object storage via Swift.
+
+          500 TiB Object Swift
+           50 TiB CephFS Manila
+
+        We hope to be able to resolve this question during work planned for 2024.
+