An ATM clustering system is a kind of workstation clusters over an ATM network. Such a system can be used as a distributed database server which requires reliable data delivery. This paper proposes an error recovery scheme at the transport layer for reliable data transfers with high throughput in the ATM clustering system. For such data transfers, acknowledgments are sent periodically as well as at the detection of packet losses. To be efficient, the periods of acknowledgments for the scheme are adaptively varied, depending on whether current transfer is either in error free phase or in an error recovery phase. The design and performance evaluation of the error recovery scheme is presented with a set of parameter values to achieve the best performance. The results show that the proposed scheme operates correctly and efficiently for the ATM clustering system. The throughput performance of the proposed scheme is superior to the SSCOP (Service Specific Connection Oriented Protocol) regardless of packet loss rates.