USB 3.2 Gen 2x2 State of the Ecosystem Review: Where Does 20Gbps USB Stand in 2020?
by Ganesh T S on October 5, 2020 10:30 AM EST- Posted in
- Storage
- SSDs
- Western Digital
- SanDisk
- ASMedia
- USB 3.2 Gen 2x2
Worst-Case Consistency, Thermals, and Power Consumption
The performance of the drives in various real-world access traces as well as synthetic workloads was brought out in the preceding sections. We also looked at the performance consistency for these cases. Power users may also be interested in performance consistency under worst-case conditions, as well as drive power consumption. The latter is also important when used with battery powered devices such as notebooks and smartphones. We analyze each of these in detail below.
Flash-based storage devices tend to slow down in unpredictable ways when subject to a large number of small-sized random writes. Many benchmarks use that scheme to pre-condition devices prior to the actual testing in order to get a worst-case representative number. Fortunately, such workloads are uncommon for direct-attached storage devices, where workloads are largely sequential in nature. Use of SLC caching as well as firmware caps to prevent overheating may cause drop in write speeds when a flash-based DAS device is subject to sustained sequential writes.
Our Sequential Writes Performance Consistency Test configures the device as a raw physical disk (after deleting configured volumes). A fio workload is set up to write sequential data to the raw drive with a block size of 128K and iodepth of 32 to cover 90% of the drive capacity. The internal temperature is recorded at either end of the workload, while the instantaneous write data rate and cumulative total write data amount are recorded at 1-second intervals.
| Sequential Write to 90% of Disk Capacity - Performance Consistency |
 |
Our primary focus is on the performance in the Haswell testbed. The Extreme PRO v2 starts off around 1800 MBps for up to 30GB of continuous writes before dropping down to around 1500 MBps for the remainder of the workload. There is a 22C rise from 56C to 78C in this process. The P50 also starts off similarly, sustaining around 1800 MBps for 13GB of continuous writes before moving down to 1400 MBps. However, the worrisome part is the drops to sub-50 MBps for 3-5 seconds at a time before recovering for around 10-15 seconds in the latter half of the workload. This symptom is seen after around 610GB of continuous writes. We do not see this problem when the drive is connected via the eGFX enclosure - however, the fill rate is around 1280 MBps consistently for the full duration of the workload in that configuration. With the USB 3.2 Gen 2 port of Alpine Ridge, the P50 is again consistent around 850 MBps. The performance of the other drives in other configurations are along expected lines, as can be observed in the above graphs.
Bus-powered devices can configure themselves to operate within the power delivery constraints of the host port. While Thunderbolt 3 ports are guaranteed to supply up to 15W for client devices, USB 3.0 ports are guaranteed to deliver only 4.5W (900mA @ 5V). In this context, it is interesting to have a fine-grained look at the power consumption profile of the various drives. Using the Plugable USBC-TKEY, the bus power consumption of the drives was tracked while processing the CrystalDiskMark workloads (separated by 30s intervals). The graphs below plot the instantaneous bus power consumption against time, while singling out the maximum and minimum power consumption numbers.
| Drive Power Consumption - CrystalDiskMark Workloads |
 |
The system lock-up during the RND4K Q32T16 workload component is evident in the above graphs, where the power consumption tracker on a different machine continues to keep tracking the power numbers while the system and the drive itself are frozen for all practical purposes. The most interesting cases for the above set are with the testbed demonstrating maximum performance - the Haswell testbed. The Extreme PRO v2 has a peak of 7.23W, and a minimum of 2.91W. The corresponding numbers for the P50 are 8.13W and 3.17W. Interestingly, we see the peak for the Extreme PRO v2 and the P50 with the eGFX configuration - 7.41W and 8.19W respectively. Using the USB SuperSpeed 10Gbps port, the drives run a bit cooler - Extreme PRO v2 peaks at 6.59W, while the P50 peaks at 7.39W. It must be noted that the P50 comes with a status LED (which glows white when the drive is connected), while the Extreme PRO v2 doesn't.
Facebook
Twitter
RSS
81 Comments
View All Comments
Eric_WVGG - Tuesday, October 6, 2020 - link
On a related note, I would love to see you guys do some kind of investigation into why we're five years into this standard and one still cannot buy an actual USB-C hub (i.e. not a port replicator).hubick - Tuesday, October 6, 2020 - link
A 3.2 hub with gen 2 ports and a 2x2 uplink would be cool!I wanted a 10gbps / gen 2 hub and got the StarTech HB31C3A1CS, which at least has a USB-C gen 2 uplink and a single USB-C gen 2 port (plus type A ports). Don't know if you can do any better than that right now.
repoman27 - Tuesday, October 6, 2020 - link
Although it's still not exactly what you're looking for, I've tried (unsuccessfully) to get people to understand what a unicorn the IOGEAR GUH3C22P is. Link: https://www.iogear.com/product/GUH3C22PIt's a 5-port USB3 10Gbps hub with a tethered USB Type-C cable on the UFP (which supports up to 85W USB PD source), two (2!) downstream facing USB Type-C ports (one of which supports up to 100W USB PD sink), and two USB Type-A ports (one of which supports up to 7.5W USB BC).
serendip - Thursday, October 8, 2020 - link
No alt mode support like for DisplayPort. I haven't found a portable type-C hub that supports DisplayPort alt mode over downstream type-C ports although some desktop docks support it.stephenbrooks - Tuesday, October 6, 2020 - link
What if I want a USB 20Gbps port on the front of my computer?Can I get a USB-C front panel and somehow connect the cable internally to the PCI-E USB card?
abufrejoval - Tuesday, October 6, 2020 - link
I am building hyperconvergent clusters for fun and for work, the home-lab one out of silent/passive 32GB RAM, 1TB SATA-SSD J5005 Atoms, the next iteration most likely from 15Watt-TDP-NUCs, an i7-10700U with 64GB RAM, 1TB NVMe SSD in testing.Clusters need short-latency, high-bandwidth interconnects, Infiniband is a classic in data centers, but NUCs offer 1Gbit Ethernet pretty much exclusively, Intel struggling to do 2.5Gbit there while Thunderbolt and USB3/4 could do much better. Only they aren’t peer-to-peer and a TB 10Gbase-T adapter sets you back further than the NUC itself, while adding lots of latency and TCP/IP, while I want RDMA.
So could we please pause for a moment and think on how we can build fabrics out of USB-X? Thunderbolt/USB4 is already about PCIe lanes, but most likely with multi-root excluded to maintain market segmentation and reduce validation effort.
I hate how the industry keeps going to 90% of something really useful and then concentrating on 200% speed instead of creating real value.
repoman27 - Wednesday, October 7, 2020 - link
Uh, Thunderbolt and USB4 are explicitly designed to support host-to-host communications already. OS / software support can be a limiting factor, but the hardware is built for it.Existing off-the-shelf solutions:
https://www.dataonstorage.com/products-solutions/k...
https://www.gosymply.com/symplyworkspace
https://www.areca.com.tw/products/thunderbolt-8050...
http://www.accusys.com.tw/T-Share/
IP over Thunderbolt is also available:
https://thunderbolttechnology.net/sites/default/fi...™%20Networking%20Bridging%20and%20Routing%20Instructional%20White%20Paper.pdf
https://support.apple.com/guide/mac-help/ip-thunde...
repoman27 - Wednesday, October 7, 2020 - link
Stupid Intel URL with ™ symbol. Let's try that again:https://thunderbolttechnology.net/sites/default/fi...
abufrejoval - Wednesday, October 7, 2020 - link
Let me tell you: You just made my day! Or more likely one or two week-ends!Not being a Mac guy, I had completely ignored Thunderbolt for a long time and never learned that it supported networking natively. From the Intel docs it looks a bit similar to Mellanox VPI and host-chaining: I can use 100Gbit links there without any switch to link three machines in a kind of “token ring” manner for Ethernet (these are hybrid adapters that would also support Infiniband, but drivers support is only support Ethernet for host-chaining). Unfortunately the effective bandwidth is only around 35GByte/s for direct hops and slows to 16GByte/s once it has to pass through another host: Not as much of an upgrade over 10Gbase-T as you’d hope for: I never really got into testing latencies, which is where the Infiniband personality of those adapters should shine.
And that’s where with TB I am hoping for significant improvements over Ethernet apart from native 40Gbit/s speed: Just right for Gluster storage!
I also used to try to get Ethernet over fiber-channel working years ago, when they were throwing out 4Gbit adapters in the data center, but even if it was specified as a standard, Ethernet over fiber never got driver support and at the higher speeds the trend went the other direction.
So I’ll definitely try to make the direct connection over TB3 work: CentOS8 should have kernel support for TB networking and the best news is that it doesn’t have to wait for TB4, but should work with TB3, too.
I’ve just seen what seemed like an incredibly cheap 4-way TB switch recommended by Anton Shilov on the TomsHardware side of this enterprise, which unfortunately is only on pre-order for now (https://eshop.macsales.com/shop/owc-thunderbolt-hu... but supposed to support TB networking. Since the NUCs are single port TB3 only, that should still do the trick and be upgradable to TB4 for just around $150… The 5Gbit USB3 Aquantia NIC wasn’t much cheaper and even the 2.5Gbit USB3 NIC are still around $40.
Exciting, exciting all that: Thank you very much for those links!
abufrejoval - Wednesday, October 7, 2020 - link
...except... I don't think that "switch" will be supporting multiple masters, same as USB.If it did, Intel would have shot themselves in the foot: 40Gbit networking on NUCs and laptops with little more than passive ables, that's almost as bad as adding a system management mode on 80486L and finding that it can be abused to implement a hypervisor (Mendel and Diane started VMware with that trick).
Yet that's exactly what consumers really thirst for.