Reiser5: Parallel scaling out on local hosts. Numbers!

From: Edward Shishkin
Date: Sun Apr 10 2022 - 12:12:43 EST


Hi all,

Earlier we announced a logical volume manager with parallel scaling out
on local hosts, which is different from the traditional RAID arrays and
is more like volume managers of networking file systems:
https://reiser4.wiki.kernel.org/index.php/Logical_Volumes_Background

Here we provide some numbers for the latest Reiser5 software release.

Note that performance of volume operations have different topicality.
E.g. performance of device removal is more topical than performance of
adding a device. Indeed, user usually wants the deleted device to go to
other needs immediately, whereas, when adding a device he gets more disk
space immediately after issuing the command (there is no need to wait
for rebalancing completion).



Hardware:

Dell OptiPlex 7050 6C2XR, Intel Core i7-7700, 16GB RAM
(4 Cores / 8 Threads)

Storage media:

DEV1: Lite-On LCS-256M6S, SSD 256GB, SATAIII, 2.5"
DEV2: Intenso 2.5" SSD TOP, 256G, SATAIII, 2.5"
DEV3: Intenso 2.5" SSD TOP, 256G, SATAIII, 2.5"
DEV4: Intenso M.2 SSD TOP, 256G, SATAIII, m.2 2280
RAM0: Block device in RAM


Software:

Reiser4-for-5.16.patch (software release 5.1.3), download at:
https://sourceforge.net/projects/reiser4/files/v5-unstable/kernel/reiser4-for-5.16.patch.gz/download


Reiser4progs-2.0.5 (software release 5.1.3), download at:
https://sourceforge.net/projects/reiser4/files/v5-unstable/progs/reiser4progs-2.0.5.tar.gz/download




Sequential RAW operations


Data set: 10G at zero offset


1. Read from RAW device

Device Speed, M/s

DEV1 470
DEV2 530


2. Write to RAW device

Device Speed, M/s


DEV1 390
DEV2 420



Sequential file operations


Stripe size: 128K
Data set: one 10G file



1. Read/Write a large file, Speed (M/s)



Nr of Disks in the Volume Write Read

1 (DEV1) 380 460
1 (DEV2) 410 518
2 (DEV1+DEV2) 695 744
3 (DEV1+DEV2+DEV3) 890 970
4 (DEV1+DEV2+DEV3+DEV4) 950 1100



2. Copy data from/to formatted device



From device To device Speed (M/s)

DEV1 DEV2 260
DEV2 DEV1 255


Volume operations


Stripe size: 128K
Data set: one 10G file


Speed of any volume operation is defined as D/T, where
D - total amount of data stored on the volume
T - operation time (including full data rebalancing/migration and sync)
Caches are dropped before each operation.

More details about logical volumes management can be found here:
https://reiser4.wiki.kernel.org/index.php/Logical_Volumes_Administration



1. Adding a device to a logical volume



Volume Device to add Speed M/s

DEV1 DEV2 284
DEV1+DEV2 DEV3 457
DEV1+DEV2+DEV3 DEV4 574


2. Removing a device from a logical volume



Volume Device to remove Speed M/s

DEV1+DEV2+DEV3+DEV4 DEV4 890
DEV1+DEV2+DEV3 DEV3 606
DEV1+DEV2 DEV2 336



3. Flushing a proxy device



More details about proxy device management can be found here:
https://reiser4.wiki.kernel.org/index.php/Proxy_Device_Administration

Before each operation all data of the logical volume are on the proxy
device. After the operation all the data are on the permanent storage
denoted as "Volume".


Volume Proxy device Speed M/s


DEV1 DEV4 228
DEV1+DEV2 DEV4 244
DEV1+DEV2+DEV3 DEV4 290


DEV1 RAM0 283
DEV1+DEV2 RAM0 301
DEV1+DEV2+DEV3 RAM0 374
DEV1+DEV2+DEV3+DEV4 RAM0 427



4. Migrating a file



More details about file migration in Reiser5 volumes can be found here:
https://reiser4.wiki.kernel.org/index.php/Transparent_File_Migration

Before each operation all data of the file are fairly distributed among
all devices-components of the logical Volume. After the operation all
file's data are stored on the Target device.


Volume Target Device Speed M/s


DEV1+DEV2+DEV3+DEV4 DEV1 387
DEV1+DEV2+DEV3 DEV1 403
DEV1+DEV2 DEV1 427


Comment. With increasing number of components, the speed of file
migration approaches (from top) the write speed to the formatted device
DEV1 (380 M/s).



Comments



Parallel O(1)-defragmentor of compound Reiser5 volumes is in plans!
Proxy devices are also a subject of defragmentation.

All file and volume operations are a subject of further performance
improvements. Currently IO requests against devices-components of a
logical volume are submitted by the same thread. This serialization
factor leads to performance drop. Once the software stability reaches
beta level, the things should be parallelized. Also currently for
simplicity _all_ data are read from the volume during its re-balancing.
It would be more reasonable to read only those data which are a subject
of migration.

Theoretical limit for the speed of adding(removing) a second device is a
double copy speed from DEV1 to DEV2 (respectively from DEV2 to DEV1).
Currently we have respectively 1.1 and 1.3 of the copy speed.