|
| 1 | +* Two questions |
| 2 | + 1. Designing a general-purpose interface between datacenters |
| 3 | + apps and programmable hardware |
| 4 | + 2. Using hardware features to better schedule low-latency datacenter |
| 5 | + applications |
| 6 | + |
| 7 | +* Intro |
| 8 | +*** Datacenter servers have increasing amounts of programmable hardware and hardware-acceleration |
| 9 | + - e.g., I/O virtualization, IOMMU, programmable NICs/flash |
| 10 | + devices, FPGAs |
| 11 | + - as a result, the hardware keeps changing and the |
| 12 | + hardware/software interface keeps changing |
| 13 | + - what does this mean for apps? |
| 14 | +*** recent work has one off solutions |
| 15 | + - Arrakis/IX for virtual I/O devices |
| 16 | + - FlexNIC, FlexTCP for programmable NICs |
| 17 | + - .. (find more related work |
| 18 | +*** but no general solution for app programmers |
| 19 | + - if an app programmer wants to make their app future proof, what |
| 20 | + do they do? |
| 21 | +*** Conclusion: We need new (lib)OS abstractions for datacenter applications |
| 22 | + - so hardware can change underneath apps, the line between |
| 23 | + hardware and software can change |
| 24 | + |
| 25 | +* Background: Why won't POSIX work? |
| 26 | + - POSIX wanted to abstract differences between devices (everything |
| 27 | + is a file) |
| 28 | + - not only has the hardware changed since we designed POSIX, the |
| 29 | + applications are very different too |
| 30 | +***** Conclusion: have new goals |
| 31 | +*** datacenter apps need to move a lot of data, not perform computation (even ML apps are limited by moving data between phases) |
| 32 | + - move memoru to network (memcached, http servers) |
| 33 | + - move ssd to network (file servers) |
| 34 | +***** Some apps don't even look at all of the data |
| 35 | +***** many things can be passed off to hardware |
| 36 | + - move to hardware hashing for checksumming |
| 37 | + - ? |
| 38 | +***** Conclusion |
| 39 | + - need to make it as cheap as possible to move data |
| 40 | + - starting to look like a middlebox? |
| 41 | +*** hardware is faster at moving data than the processor |
| 42 | + - I/O devices can now move data faster than processor, so we need |
| 43 | + to have a zero-copy interface |
| 44 | + - POSIX is fundamentally not: it's built around copying into app |
| 45 | + memory and then out to device |
| 46 | + - we need an interface that is able to hand pre-allocated buffers |
| 47 | + to app and have to app hand off buffer to another device, |
| 48 | + potentially without looking at some or any of the buffer |
| 49 | + |
| 50 | +* Zero-copy event queues |
| 51 | + - replaces socket, file descriptor abstractions |
| 52 | + - open(), accept() returns an event queue (id) |
| 53 | + - has concept of granularity (not just a stream of data) |
| 54 | + - moves data with pointers, not by streaming into a buffer |
| 55 | + - use COW every time the pointer is transferred to another address |
| 56 | + space to avoid complex pointer hand-offs |
| 57 | + - use user-level page tables and a directed TLB shoot-down to |
| 58 | + reduce cost of setting COW |
| 59 | +***** Interface |
| 60 | + - qid = open(file) |
| 61 | + - qid = listen() |
| 62 | + - qid = accept(qid) |
| 63 | + - insert(qid, scatter gather array) |
| 64 | + - *sga = head(qid) |
| 65 | + - *sga = dequeue(qid) |
| 66 | + - filter(qid, *filter_func) |
| 67 | + - merge(qid, qid) |
| 68 | + - sort(qid, *sort_func) |
| 69 | +* Benefits |
| 70 | +*** No copying latency (at least 2K cycles for a 4K page) |
| 71 | +*** Less cache pollution |
| 72 | + - Only data that the app needs has to be brought into cache |
| 73 | +*** can be implemented in hardware or software or both |
| 74 | + - even advanced filtering, merging and sorting can be implemented |
| 75 | + in hardware easily |
| 76 | + |
| 77 | +* Datacenters have increasingly demanding workloads (low latency, low tail latency) |
| 78 | + - driving much of the programmable and hardware acceleration |
| 79 | + - how can we use this hardware for these workloads |
| 80 | +*** Current solution: Datacenters do not make good use of cores for these apps |
| 81 | + - context switches are expensive and increase tail latency, so they pin apps to cores |
| 82 | + - interrupts are expensive and increase tail latency, so they poll |
| 83 | + - both are terrible for CPU utilization |
| 84 | +*** Key Observation: datacenter apps are event-based programs, not long running serial programs |
| 85 | + - interrupt scheduling is ineffective for datacenter workloads |
| 86 | + when they have natural yield points |
| 87 | + - polling helps but takes too much time to switch back, so only |
| 88 | + works for low latency workloads if they are pinned |
| 89 | + |
| 90 | +* Cooperative Event Scheduling |
| 91 | +*** Idea |
| 92 | + - yield between every event to check for higher priority tasks |
| 93 | + - process with low latency and go back to lower priority processes |
| 94 | +*** Design requirements |
| 95 | + - scheduling decisions must be fast |
| 96 | + - context switches must be cheap |
| 97 | +*** Possible implementations |
| 98 | + - move scheduling into hardware based on queues (IOCPU instead of |
| 99 | + IOMMU?) |
| 100 | + - tagged TLBs and partitioned caches kept warm for low latency |
| 101 | + apps |
| 102 | + - yielding between events means that old cached data might not be |
| 103 | + useful for next event anyway (experiment: flush cache between |
| 104 | + every libevent/memcached handler and check performance) |
| 105 | + |
| 106 | +* Summary |
| 107 | + - we can't keep changing the hardware without some abstractions to |
| 108 | + buffer apps from those changes |
| 109 | + - we can't effectively schedule low latency apps without co-design |
| 110 | + between the app, the OS and the hardware |
0 commit comments