|
| No | With |
| |
| Lookaside | Lookaside | Win | ||
100 Mb/s | 173 (9) | 103 (3.9) | 40.1% | ||
10 Mb/s | 370 | (14) | 163 (2.9) | 55.9% | |
1 Mb/s | 2688 | (39) | 899 (26.4) | 66.6% | |
100 Kb/s | 30531 (1490) | 8567 (463.9) | 71.9% | ||
|
|
|
|
| |
This table gives the total operation latency (in seconds) for the CDA benchmark of Section 5.2 at different bandwidths, with and without lookaside to a
Figure 11. Off-machine Lookaside
can be used for lookaside. Distributed hash tables (DHTs) are one such source. There is growing interest in DHTs such as Pastry [16], Chord [20], Tapestry [23] and CAN [15]. There is also growing interest in
Lookaside caching enables a conventional dis- tributed file system based on the
We have recently extended the prototype imple- mentation described in Section 4 to support off- machine CAS providers. Experiments with this ex- tended prototype confirm its performance benefits. For the ISR benchmark described in Section 5.2, Fig- ure 11 shows the performance benefit of using a LAN- attached CAS provider with same contents as the DVD of Figure 8. Since the CAS provider is on a faster ma- chine than the file server, Figure 11 shows a substantial benefit even at 100 Mb/s.
Another potential application of lookaside caching is in implementing a form of cooperative caching [1, 4]. A collection of distributed file system clients with mutual trust (typically at one location) can ex- port each other’s file caches as CAS providers. No protocol is needed to maintain mutual cache consis- tency; divergent caches may, at worst, reduce looka-
side performance improvement. This form of cooper- ative caching can be especially valuable in situations where the clients have LAN connectivity to each other, but poor connectivity to a distant file server. The heavy price of a cache miss on a large file is then borne only by the first client to access the file. Misses elsewhere are serviced at LAN speeds, provided the file has not been replaced in the first client’s cache.
7Conclusion
“Sneakernet,” the informal term for manual trans- port of data, is alive and well today in spite of advances in networking and distributed file systems. Early in this paper, we examined why this is the case. Carrying your data on a portable storage device gives you full confi- dence that you will be to access that data anywhere, regardless of network quality, network or server out- ages, and machine configuration. Unfortunately, this confidence comes at a high price. Remembering to carry the right device, ensuring that data on it is cur- rent, tracking updates by collaborators, and guarding against loss, theft and damage are all burdens borne by the user. Most harried mobile users would gladly del- egate these chores if only they could be confident that they would have good access to their critical data at all times and places.
Lookaside caching suggests a way of achieving this goal. Let the true home of your data be in a distributed file system. Make a copy of your critical data on a portable storage device. If you find yourself needing to access the data in a desperate situation, just use the device directly — you are no worse off than if you were relying on sneakernet. In all other situations, use the device for lookaside caching. On a slow net- work or with a heavily loaded server, you will benefit from improved performance. With network or server outages, you will benefit from improved availability if your distributed file system supports disconnected op- eration and if you have hoarded all your
Notice that you make the decision to use the de- vice directly or via lookaside caching at the point of use, not a priori. This preserves maximum flexibility up front, when there may be uncertainty about the ex- act future locations where you will need to access the data. Lookaside caching thus integrates portable stor- age devices and distributed file systems in a manner that combines their strengths. It preserves the intrinsic advantages of performance, availability and ubiquity
10