Academia.eduAcademia.edu

Figure 8 – uploaded by Mark S Manasse

See full PDFdownloadDownload figure
Fig 9. Speedup based on a sample pair of PPT files, as a function of horizon size (sampled between 40 and 2000) and recursion level (between | and 4). In the next set of experiments, we first tune the chunking parameters by choosing a horizon and window size h and w, respectively. We examine the impact of recursion next, and evaluate the bandwidth savings against RSYNC and the local diff utilities. The data set used for these experiments consists

Figure 9 Speedup based on a sample pair of PPT files, as a function of horizon size (sampled between 40 and 2000) and recursion level (between | and 4). In the next set of experiments, we first tune the chunking parameters by choosing a horizon and window size h and w, respectively. We examine the impact of recursion next, and evaluate the bandwidth savings against RSYNC and the local diff utilities. The data set used for these experiments consists

Related Figures (13)

Fig. 1. The basic RDC protocol in LBFS [24] starts with both client and server machines dividing their respective files intc chunks and computing signatures for each chunk. The server then communicates the list of signatures to the client, which checks it against its own to determine which chunks it needs to request from the server. Upon receipt of the missing chunks, the clien combines the received chunks with its own seed ones to reassemble a copy of the new file Fs.
Fig. 1. The basic RDC protocol in LBFS [24] starts with both client and server machines dividing their respective files intc chunks and computing signatures for each chunk. The server then communicates the list of signatures to the client, which checks it against its own to determine which chunks it needs to request from the server. Upon receipt of the missing chunks, the clien combines the received chunks with its own seed ones to reassemble a copy of the new file Fs.
Fig. 3. Calculated similarity curves for (b=6,f=16) showing the detection probability of finding 1, 2, ..., 16 matching traits, given the actual fraction of matching chunk signatures. The bold curve is the one used in DFSR, corresponding to 5 matching traits.
Fig. 3. Calculated similarity curves for (b=6,f=16) showing the detection probability of finding 1, 2, ..., 16 matching traits, given the actual fraction of matching chunk signatures. The bold curve is the one used in DFSR, corresponding to 5 matching traits.
Fig. 4. Example of misaligned cut points for m=3 and h=a when using the point filter method, where m is the minimum chunk size and h is the value at which cut-points are chosen.
Fig. 4. Example of misaligned cut points for m=3 and h=a when using the point filter method, where m is the minimum chunk size and h is the value at which cut-points are chosen.
Fig. 5. Replicating a file F from a server to a client involves marshaling the file F and its metadata into a staged representation M(F) on the server, negotiating the traits and the RDC recursion level, identifying a set of similar staged files M(F,), M(F2),... on the client, using RDC to transfer M(F), and finally un-marshaling the transferred M(F) into a copy of the file F on the client.
Fig. 5. Replicating a file F from a server to a client involves marshaling the file F and its metadata into a staged representation M(F) on the server, negotiating the traits and the RDC recursion level, identifying a set of similar staged files M(F,), M(F2),... on the client, using RDC to transfer M(F), and finally un-marshaling the transferred M(F) into a copy of the file F on the client.
Fig. 6. The staged representation of a file F contains the marshaled file metadata and data streams M(F) stored in compressed form CM(F), the RDC signatures for M(F), and the jump table for CM(F). The jump table is used for seeking inside CM(F) given logical offsets into M(F), and is based on storing the offsets in CM(F) for every 8K segment in M(F); seeking inside a segment requires decompression of the segment.
Fig. 6. The staged representation of a file F contains the marshaled file metadata and data streams M(F) stored in compressed form CM(F), the RDC signatures for M(F), and the jump table for CM(F). The jump table is used for seeking inside CM(F) given logical offsets into M(F), and is based on storing the offsets in CM(F) for every 8K segment in M(F); seeking inside a segment requires decompression of the segment.
Fig. 7. RDC speedup factors without recursion, using various combinations of the horizon / and window size w, compared to RSYNC. The average chunk size is C=2h+1.
Fig. 7. RDC speedup factors without recursion, using various combinations of the horizon / and window size w, compared to RSYNC. The average chunk size is C=2h+1.
Fig. 8. RDC speedup factors for a fixed horizon and window size (hA=256, w=32) but with variable levels of recursion, compared to RSYNC. The biggest boost is achieved for large files.
Fig. 8. RDC speedup factors for a fixed horizon and window size (hA=256, w=32) but with variable levels of recursion, compared to RSYNC. The biggest boost is achieved for large files.
Table. 11. Speedup factors for RDC vs. RSYNC for large files. Fig. 9 summarizes an experiment performed on a PPT file selected at random. A combination of horizon size between 80 and 160 and 4 recursive applications of RDC provide the highest speedup (about 13x), while recursion has little effect for the higher horizon sizes.
Table. 11. Speedup factors for RDC vs. RSYNC for large files. Fig. 9 summarizes an experiment performed on a PPT file selected at random. A combination of horizon size between 80 and 160 and 4 recursive applications of RDC provide the highest speedup (about 13x), while recursion has little effect for the higher horizon sizes.
Fig. 10. Speedup of recursive RDC and RSYNC normalizec to xdelta. Note these figures all include compression.
Fig. 10. Speedup of recursive RDC and RSYNC normalizec to xdelta. Note these figures all include compression.
Table 13. Breakdown of RDC and non-RDC transfers in a deployment scenario.
Table 13. Breakdown of RDC and non-RDC transfers in a deployment scenario.
Fig. 12. Comparison of the transfer bandwidth required when using similarity detection with different parameters and RDC (without recursion) vs. LBFS.
Fig. 12. Comparison of the transfer bandwidth required when using similarity detection with different parameters and RDC (without recursion) vs. LBFS.
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved