James Bottomley wrote:
The two target architectures perform essentially identical functions, so
there's only really room for one in the kernel. Right at the moment,
it's STGT. Problems in STGT come from the user<->kernel boundary which
can be mitigated in a variety of ways. The fact that the figures are
pretty much comparable on non IB networks shows this.
I really need a whole lot more evidence than at worst a 20% performance
difference on IB to pull one implementation out and replace it with
another. Particularly as there's no real evidence that STGT can't be
tweaked to recover the 20% even on IB.
James,
Although the performance difference between STGT and SCST is apparent, this isn't the only point why SCST is better. I've already written about it many times in various mailing lists, but let me summarize it one more time here.
As you know, almost all kernel parts can be done in user space, including all the drivers, networking, I/O management with block/SCSI initiator subsystem and disk cache manager. But does it mean that currently Linux kernel is bad and all the above should be (re)done in user space instead? I believe, not. Linux isn't a microkernel for very pragmatic reasons: simplicity and performance. So, additional important point why SCST is better is simplicity.
For SCSI target, especially with hardware target card, data are came from kernel and eventually served by kernel, which does actual I/O or getting/putting data from/to cache. Dividing requests processing between user and kernel space creates unnecessary interface layer(s) and effectively makes the requests processing job distributed with all its complexity and reliability problems. From my point of view, having such distribution, where user space is master side and kernel is slave is rather wrong, because:
1. It makes kernel depend from user program, which services it and provides for it its routines, while the regular paradigm is the opposite: kernel services user space applications. As a direct consequence from it that there is no real protection for the kernel from faults in the STGT core code without excessive effort, which, no surprise, wasn't currently done and, seems, is never going to be done. So, on practice debugging and developing under STGT isn't easier, than if the whole code was in the kernel space, but, actually, harder (see below why).
2. It requires new complicated interface between kernel and user spaces that creates additional maintenance and debugging headaches, which don't exist for kernel only code. Linus Torvalds some time ago perfectly described why it is bad, see http://lkml.org/lkml/2007/4/24/451, http://lkml.org/lkml/2006/7/1/41 and http://lkml.org/lkml/2007/4/24/364.
3. It makes for SCSI target impossible to use (at least, on a simple and sane way) many effective optimizations: zero-copy cached I/O, more control over read-ahead, device queue unplugging-plugging, etc. One example of already implemented such features is zero-copy network data transmission, done in simple 260 lines put_page_callback patch. This optimization is especially important for the user space gate (scst_user module), see below for details.
The whole point that development for kernel is harder, than for user space, is totally nonsense nowadays. It's different, yes, in some ways more limited, yes, but not harder. For ones who need gdb (I for many years - don't) kernel has kgdb, plus it also has many not available for user space or more limited there debug facilities like lockdep, lockup detection, oprofile, etc. (I don't mention wider choice of more effectively implemented synchronization primitives and not only them).
For people who need complicated target devices emulation, like, e.g., in case of VTL (Virtual Tape Library), where there is a need to operate with large mmap'ed memory areas, SCST provides gateway to the user space (scst_user module), but, in contrast with STGT, it's done in regular "kernel - master, user application - slave" paradigm, so it's reliable and no fault in user space device emulator can break kernel and other user space applications. Plus, since SCSI target state machine and memory management are in the kernel, it's very effective and allows only one kernel-user space switch per SCSI command.
Also, I should note here, that in the current state STGT in many aspects doesn't fully conform SCSI specifications, especially in area of management events, like Unit Attentions generation and processing, and it doesn't look like somebody cares about it. At the same time, SCST pays big attention to fully conform SCSI specifications, because price of non-conformance is a possible user's data corruption.
Returning to performance, modern SCSI transports, e.g. InfiniBand, have as low link latency as 1(!) microsecond. For comparison, the inter-thread context switch time on a modern system is about the same, syscall time - about 0.1 microsecond. So, only ten empty syscalls or one context switch add the same latency as the link. Even 1Gbps Ethernet has less, than 100 microseconds of round-trip latency.
You, probably, know, that QLogic Fibre Channel target driver for SCST allows commands being executed either directly from soft IRQ, or from the corresponding thread. There is a steady 5-7% difference in IOPS between those modes on 512 bytes reads on nullio using 4Gbps link. So, a single additional inter-kernel-thread context switch costs 5-7% of IOPS.
Another source of additional unavoidable with the user space approach latency is data copy to/from cache. With the fully kernel space approach, cache can be used directly, so no extra copy will be needed. We can estimate how much latency the data copying adds. On the modern systems memory copy throughput is less than 2GB/s, so on 20Gbps InfiniBand link it almost doubles data transfer latency.
So, putting code in the user space you should accept the extra latency it adds. Many, if not most, real-life workloads more or less latency, not throughput, bound, so there shouldn't be surprise that single stream "dd if=/dev/sdX of=/dev/null" on initiator gives too low values. Such "benchmark" isn't less important and practical, than all the multithreaded latency insensitive benchmarks, which people like running, because it does essentially the same as most Linux processes do when they read data from files.
You may object me that the target's backstorage device(s) latency is a lot more, than 1 microsecond, but that is relevant only if data are read/written from/to the actual backstorage media, not from the cache, even from the backstorage device's cache. Nothing prevents target from having 8 or even 64GB of cache, so most even random accesses could be served by it. This is especially important for sync writes.
Thus, why SCST is better:
1. It is more simple, because it's monolithic, so all its components are in one place and communicate using direct function calls. Hence, it is smaller, faster, more reliable and maintainable. Currently it's bigger, than STGT, just because it supports more features, see (2).
2. It supports more features: 1 to many pass-through support with all necessary for it functionality, including support for non-disk SCSI devices, like tapes, SGV cache, BLOCKIO, where requests converted to bio's and directly sent to block level (this mode is effective for random mostly workloads with data set size >> memory size on the target), etc.
3. It has better performance and going to have it even better. SCST only now enters in the phase, where it starts exploiting all advantages of being in the kernel. Particularly, zero-copy cached I/O is currently being implemented.
4. It provides safer and more effective interface to emulate target devices in the user space via scst_user module.
5. It much more confirms to SCSI specifications (see above).