Ethereum

    RowDAS prototype cuts CPU work

    行政By No Comments3 Mins Read

    An Ethereum prototype that divides blob-recovery duties among nodes reported an 11–18× reduction in estimated reconstruction computing work across 1,000-node simulations. The results suggest operators could reduce duplicated work through a smaller change than the full RowDAS networking proposal.

    Researcher Csaba Kiraly’s Sept. 3 report describes the reduced design as a possible first step toward RowDAS. It assigns recovery duties without introducing the new row-networking channels in the full proposal.

    Blobs carry data used by layer-2 rollups. PeerDAS, Ethereum’s system for checking that blob data is available, lets nodes download only part of it. High-custody nodes hold at least 64 of the 128 data columns, enough to rebuild missing blob data; supernodes hold all 128.

    Related Reading

    Ethereum’s data bloat threatens home staking amid surge toward 1.2TB need

    Many high-custody nodes can repeat the same reconstruction. The reduced design assigns particular blobs to them first, allowing others to receive the recovered data instead of immediately rebuilding it themselves.

    What the Ethereum blob-recovery simulations show

    In one configuration with four blobs, 10% supernodes and no columns withheld, the estimated network-wide reconstruction cost fell from 48.6 CPU-seconds under the PeerDAS model to 2.75 CPU-seconds under the reduced design. At a 20% supernode share, the corresponding figures were 91 and 6.6 CPU-seconds.

    Those totals describe accumulated computing work across the simulated network, rather than elapsed recovery time. The accounting applies a measured 162-millisecond cost per blob recovery on a Ryzen 9 8945HS processor. Transaction speeds and fee savings were outside the reported measurements.

    Reported simulation CPU work for 1,000 nodes and four blobs with no columns withheld: 48.6 versus 2.75 CPU-seconds at 10% supernodes and 91 versus 6.6 at 20%, comparing PeerDAS with the reduced RowDAS variant. CPU work is not elapsed time; high-custody nodes are still required and the report had no devnet results.Reported simulation CPU work for 1,000 nodes and four blobs with no columns withheld: 48.6 versus 2.75 CPU-seconds at 10% supernodes and 91 versus 6.6 at 20%, comparing PeerDAS with the reduced RowDAS variant. CPU work is not elapsed time; high-custody nodes are still required and the report had no devnet results.

    The PeerDAS baseline already includes randomized waiting and checks that suppress duplicate reconstruction. The comparison therefore gives existing client behavior credit for the work those delays save.